Listing all models.">

List models

GET /v1/models

Lists all language models that are available for use with the service. The information includes the name of the model and its minimum sampling rate in Hertz, among other things. The ordering of the list of models can change from call to call; do not rely on an alphabetized or static list of models.

See also: Listing all models.

Request

Response

Response message.

warnings

array

Warnings associated with the error.

Listing a specific model.">

Get a model

GET /v1/models/{model_id}

Gets information for a single specified language model that is available for use with the service. The information includes the name of the model and its minimum sampling rate in Hertz, among other things.

See also: Listing a specific model.

Request

Path Parameters

model_id

Required *

string

The identifier of the model in the form of its name from the output of the List models method.

Query Parameters

Response

200 - OK. The request succeeded.

name

string

The name of the model for use as an identifier in calls to the service (for example, `en-US_BroadbandModel`).

language

Warnings associated with the error.

Making a basic HTTP request.

Streaming mode

For requests to transcribe live audio as it becomes available, you must set the `Transfer-Encoding` header to `chunked` to use streaming mode. In streaming mode, the service closes the connection (status code 408) if it does not receive at least 15 seconds of audio (including silence) in any 30-second period. The service also closes the connection (status code 400) if it detects no speech for `inactivity_timeout` seconds of streaming audio; use the `inactivity_timeout` parameter to change the default of 30 seconds.

See also:

Audio formats (content types)

The service accepts audio in the following formats (MIME types).

For formats that are labeled Required, you must use the `Content-Type` header with the request to specify the format of the audio.
For all other formats, you can omit the `Content-Type` header or specify `application/octet-stream` with the header to have the service automatically detect the format of the audio. (With the `curl` command, you can specify either `"Content-Type:"` or `"Content-Type: application/octet-stream"`.)

Where indicated, the format that you specify must include the sampling rate and can optionally include the number of channels and the endianness of the audio.

`audio/alaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/basic` (Required. Use only with narrowband models.)
`audio/flac`
`audio/g729` (Use only with narrowband models.)
`audio/l16` (Required. Specify the sampling rate (`rate`) and optionally the number of channels (`channels`) and endianness (`endianness`) of the audio.)
`audio/mp3`
`audio/mpeg`
`audio/mulaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/ogg` (The service automatically detects the codec of the input audio.)
`audio/ogg;codecs=opus`
`audio/ogg;codecs=vorbis`
`audio/wav` (Provide audio with a maximum of nine channels.)
`audio/webm` (The service automatically detects the codec of the input audio.)
`audio/webm;codecs=opus`
`audio/webm;codecs=vorbis`

The sampling rate of the audio must match the sampling rate of the model for the recognition request: for broadband models, at least 16 kHz; for narrowband models, at least 8 kHz. If the sampling rate of the audio is higher than the minimum required rate, the service down-samples the audio to the appropriate rate. If the sampling rate of the audio is lower than the minimum required rate, the request fails.

See also: Supported audio formats.

Large speech models and Next-generation models

The service supports large speech models and next-generation `Multimedia` (16 kHz) and `Telephony` (8 kHz) models for many languages. Large speech models and next-generation models have higher throughput than the service's previous generation of `Broadband` and `Narrowband` models. When you use large speech models and next-generation models, the service can return transcriptions more quickly and also provide noticeably better transcription accuracy.

You specify a large speech model or next-generation model by using the `model` query parameter, as you do a previous-generation model. Only the next-generation models support the `low_latency` parameter, and all large speech models and next-generation models support the `character_insertion_bias` parameter. These parameters are not available with previous-generation models.

Large speech models and next-generation models do not support all of the speech recognition parameters that are available for use with previous-generation models. Next-generation models do not support the following parameters:

`acoustic_customization_id`
`keywords` and `keywords_threshold`
`processing_metrics` and `processing_metrics_interval`
`word_alternatives_threshold`

Important: Effective 31 July 2023, all previous-generation models will be removed from the service and the documentation. Most previous-generation models were deprecated on 15 March 2022. You must migrate to the equivalent large speech model or next-generation model by 31 July 2023. For more information, see Migrating to large speech models.

See also:

Multipart speech recognition

Note: The asynchronous HTTP interface, WebSocket interface, and Watson SDKs do not support multipart speech recognition.

The HTTP `POST` method of the service also supports multipart speech recognition. With multipart requests, you pass all audio data as multipart form data. You specify some parameters as request headers and query parameters, but you pass JSON metadata as form data to control most aspects of the transcription. You can use multipart recognition to pass multiple audio files with a single request.

Use the multipart approach with browsers for which JavaScript is disabled or when the parameters used with the request are greater than the 8 KB limit imposed by most HTTP servers and proxies. You can encounter this limit, for example, if you want to spot a very large number of keywords.

See also: Making a multipart HTTP request.">

Recognize audio

POST /v1/recognize

Sends audio and returns transcription results for a recognition request. You can pass a maximum of 100 MB and a minimum of 100 bytes of audio with a request. The service automatically detects the endianness of the incoming audio and, for audio that includes multiple channels, downmixes the audio to one-channel mono during transcoding. The method returns only final results; to enable interim results, use the WebSocket API. (With the `curl` command, use the `--data-binary` option to upload the file for the request.)

See also: Making a basic HTTP request.

Streaming mode

See also:

Audio formats (content types)

The service accepts audio in the following formats (MIME types).

For formats that are labeled Required, you must use the `Content-Type` header with the request to specify the format of the audio.
For all other formats, you can omit the `Content-Type` header or specify `application/octet-stream` with the header to have the service automatically detect the format of the audio. (With the `curl` command, you can specify either `"Content-Type:"` or `"Content-Type: application/octet-stream"`.)

Where indicated, the format that you specify must include the sampling rate and can optionally include the number of channels and the endianness of the audio.

`audio/alaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/basic` (Required. Use only with narrowband models.)
`audio/flac`
`audio/g729` (Use only with narrowband models.)
`audio/l16` (Required. Specify the sampling rate (`rate`) and optionally the number of channels (`channels`) and endianness (`endianness`) of the audio.)
`audio/mp3`
`audio/mpeg`
`audio/mulaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/ogg` (The service automatically detects the codec of the input audio.)
`audio/ogg;codecs=opus`
`audio/ogg;codecs=vorbis`
`audio/wav` (Provide audio with a maximum of nine channels.)
`audio/webm` (The service automatically detects the codec of the input audio.)
`audio/webm;codecs=opus`
`audio/webm;codecs=vorbis`

See also: Supported audio formats.

Large speech models and Next-generation models

`acoustic_customization_id`
`keywords` and `keywords_threshold`
`processing_metrics` and `processing_metrics_interval`
`word_alternatives_threshold`

See also:

Multipart speech recognition

Note: The asynchronous HTTP interface, WebSocket interface, and Watson SDKs do not support multipart speech recognition.

See also: Making a multipart HTTP request.

Request

Request Body

Path Parameters

Query Parameters

model

string

array

An array of keyword strings to spot in the audio. Each keyword string can include one or more string tokens. Keywords are spotted only in the final results, not in interim hypotheses. If you specify any keywords, you must also specify a keywords threshold. Omit the parameter or specify an empty array if you do not need to spot keywords.

You can spot a maximum of 1000 keywords with a single request. A single keyword can have a maximum length of 1024 characters, though the maximum effective length for double-byte languages might be shorter. Keywords are case-insensitive.

See Keyword spotting.

keywords_threshold

number

A confidence value that is the lower bound for spotting a keyword. A word is considered to match a keyword if its confidence is greater than or equal to the threshold. Specify a probability between 0.0 and 1.0. If you specify a threshold, you must also specify one or more keywords. The service performs no keyword spotting if you omit either parameter. See Keyword spotting.

max_alternatives

integer

The maximum number of alternative transcripts that the service is to return. By default, the service returns a single transcript. If you specify a value of `0`, the service uses the default value, `1`. See Maximum alternatives.

word_alternatives_threshold

number

A confidence value that is the lower bound for identifying a hypothesis as a possible word alternative (also known as "Confusion Networks"). An alternative word is considered if its confidence is greater than or equal to the threshold. Specify a probability between 0.0 and 1.0. By default, the service computes no alternative words. See Word alternatives.

word_confidence

boolean

If `true`, the service returns a confidence measure in the range of 0.0 to 1.0 for each word. By default, the service returns no word confidence scores. See Word confidence.

timestamps

boolean

If `true`, the service returns time alignment for each word. By default, no timestamps are returned. See Word timestamps.

profanity_filter

boolean

If `true`, the service filters profanity from all output except for keyword results by replacing inappropriate words with a series of asterisks. Set the parameter to `false` to return results with no censoring.

Note: The parameter can be used with US English and Japanese transcription only. See Profanity filtering.

smart_formatting

boolean

If `true`, the service converts dates, times, series of digits and numbers, phone numbers, currency values, and internet addresses into more readable, conventional representations in the final transcript of a recognition request. For US English, the service also converts certain keyword strings to punctuation symbols. By default, the service performs no smart formatting.

Note: The parameter can be used with US English, Japanese, and Spanish (all dialects) transcription only.

See Smart formatting.

smart_formatting_version

integer

Smart formatting version for large speech models and next-generation models is supported in US English, Brazilian Portuguese, French, German, Spanish and French Canadian languages.

speaker_labels

boolean

If `true`, the response includes labels that identify which words were spoken by which participants in a multi-person exchange. By default, the service returns no speaker labels. Setting `speaker_labels` to `true` forces the `timestamps` parameter to be `true`, regardless of whether you specify `false` for the parameter.

For previous-generation models, the parameter can be used with Australian English, US English, German, Japanese, Korean, and Spanish (both broadband and narrowband models) and UK English (narrowband model) transcription only.
For large speech models and next-generation models, the parameter can be used with all available languages.

See Speaker labels.

grammar_name

string

The name of a grammar that is to be used with the recognition request. If you specify a grammar, you must also use the `language_customization_id` parameter to specify the name of the custom language model for which the grammar is defined. The service recognizes only strings that are recognized by the specified grammar; it does not recognize other custom words from the model's words resource.

See Using a grammar for speech recognition.

redaction

boolean

If `true`, the service redacts, or masks, numeric data from final transcripts. The feature redacts any number that has three or more consecutive digits by replacing each digit with an `X` character. It is intended to redact sensitive numeric data, such as credit card numbers. By default, the service performs no redaction.

When you enable redaction, the service automatically enables smart formatting, regardless of whether you explicitly disable that feature. To ensure maximum security, the service also disables keyword spotting (ignores the `keywords` and `keywords_threshold` parameters) and returns only a single final transcript (forces the `max_alternatives` parameter to be `1`).

Note: The parameter can be used with US English, Japanese, and Korean transcription only.

See Numeric redaction.

audio_metrics

boolean

If `true`, requests detailed information about the signal characteristics of the input audio. The service returns audio metrics with the final transcription results. By default, the service returns no audio metrics.

See Audio metrics.

end_of_phrase_silence_time

number

Specifies the duration of the pause interval at which the service splits a transcript into multiple final results. If the service detects pauses or extended silence before it reaches the end of the audio stream, its response can include multiple final results. Silence indicates a point at which the speaker pauses between spoken words or phrases.

Specify a value for the pause interval in the range of 0.0 to 120.0.

A value greater than 0 specifies the interval that the service is to use for speech recognition.
A value of 0 indicates that the service is to use the default interval. It is equivalent to omitting the parameter.

The default pause interval for most languages is 0.8 seconds; the default for Chinese is 0.6 seconds.

See End of phrase silence time.

split_transcript_at_phrase_end

boolean

If `true`, directs the service to split the transcript into multiple final results based on semantic features of the input, for example, at the conclusion of meaningful phrases such as sentences. The service bases its understanding of semantic features on the base language model that you use with a request. Custom language models and grammars can also influence how and where the service splits a transcript.

By default, the service splits transcripts based solely on the pause interval. If the parameters are used together on the same request, `end_of_phrase_silence_time` has precedence over `split_transcript_at_phrase_end`.

See Split transcript at phrase end.

speech_detector_sensitivity

number

The sensitivity of speech activity detection that the service is to perform. Use the parameter to suppress word insertions from music, coughing, and other non-speech events. The service biases the audio it passes for speech recognition by evaluating the input audio against prior models of speech and non-speech activity.

Specify a value between 0.0 and 1.0:

0.0 suppresses all audio (no speech is transcribed).
0.5 (the default) provides a reasonable compromise for the level of sensitivity.
1.0 suppresses no audio (speech detection sensitivity is disabled).

The values increase on a monotonic curve. Specifying one or two decimal places of precision (for example, `0.55`) is typically more than sufficient.

The parameter is supported with all large speech models, next-generation models and with most previous-generation models. See Speech detector sensitivity and Language model support.

sad_module

integer

Detects speech boundaries within the audio stream with better performance, improved noise suppression, faster responsiveness, and increased accuracy.

Specify `sad_module: 2`

See Speech Activity Detection (SAD).

background_audio_suppression

number

The level to which the service is to suppress background audio based on its volume to prevent it from being transcribed as speech. Use the parameter to suppress side conversations or background noise.

Specify a value in the range of 0.0 to 1.0:

0.0 (the default) provides no suppression (background audio suppression is disabled).
0.5 provides a reasonable level of audio suppression for general usage.
1.0 suppresses all audio (no audio is transcribed).

The values increase on a monotonic curve. Specifying one or two decimal places of precision (for example, `0.55`) is typically more than sufficient.

The parameter is supported with all large speech models, next-generation models and with most previous-generation models. See Background audio suppression and Language model support.

low_latency

boolean

If `true` for next-generation `Multimedia` and `Telephony` models that support low latency, directs the service to produce results even more quickly than it usually does. Next-generation models produce transcription results faster than previous-generation models. The `low_latency` parameter causes the models to produce results even more quickly, though the results might be less accurate when the parameter is used.

The parameter is not available for large speech models and previous-generation `Broadband` and `Narrowband` models. It is available for most next-generation models.

For a list of next-generation models that support low latency, see Supported next-generation language models.
For more information about the `low_latency` parameter, see Low latency.

character_insertion_bias

number

For large speech models and next-generation models, an indication of whether the service is biased to recognize shorter or longer strings of characters when developing transcription hypotheses. By default, the service is optimized to produce the best balance of strings of different lengths.

The default bias is 0.0. The allowable range of values is -1.0 to 1.0.

Negative values bias the service to favor hypotheses with shorter strings of characters.
Positive values bias the service to favor hypotheses with longer strings of characters.

As the value approaches -1.0 or 1.0, the impact of the parameter becomes more pronounced. To determine the most effective value for your scenario, start by setting the value of the parameter to a small increment, such as -0.1, -0.05, 0.05, or 0.1, and assess how the value impacts the transcription results. Then experiment with different values as necessary, adjusting the value by small increments.

The parameter is not available for previous-generation models.

See Character insertion bias.

Response

200 - OK. The request succeeded.

results

array

An array of `SpeechRecognitionResult` objects that can include interim and final results (interim results are returned only if supported by the method). Final results are guaranteed not to change; interim results might be replaced by further interim results and eventually final results.

For the HTTP interfaces, all results arrive at the same time. For the WebSocket interface, results can be sent as multiple separate responses. The service periodically sends updates to the results list. The `result_index` is incremented to the lowest index in the array that has changed for new results.

For more information, see Understanding speech recognition results.

result_index

integer

An index that indicates a change point in the `results` array. The service increments the index for additional results that it sends for new audio for the same request. All results with the same index are delivered at the same time. The same index can include multiple final results that are delivered with the same response.

speaker_labels

array

An array of `SpeakerLabelsResult` objects that identifies which words were spoken by which speakers in a multi-person exchange. The array is returned only if the `speaker_labels` parameter is `true`. When interim results are also requested for methods that support them, it is possible for a `SpeechRecognitionResults` object to include only the `speaker_labels` field.

processing_metrics

ProcessingMetrics

If processing metrics are requested, information about the service's processing of the input audio. Processing metrics are not available with the synchronous Recognize audio method.

audio_metrics

AudioMetrics

If audio metrics are requested, information about the signal characteristics of the input audio.

warnings

array

An array of warning messages associated with the request:

Warnings for invalid parameters or fields can include a descriptive message and a list of invalid argument strings, for example, `"Unknown arguments:"` or `"Unknown url query arguments:"` followed by a list of the form `"{invalid_arg_1}, {invalid_arg_2}."` (If you use the `character_insertion_bias` parameter with a previous-generation model, the warning message refers to the parameter as `lambdaBias`.)
The following warning is returned if the request passes a custom model that is based on an older version of a base model for which an updated version is available: `"Using previous version of base model, because your custom model has been built with it. Please note that this version will be supported only for a limited time. Consider updating your custom model to the new base model. If you do not do that you will be automatically switched to base model when you used the non-updated custom model."`

In both cases, the request succeeds despite the warnings.

enriched_results

EnrichedResults

If enriched results are requested, transcription with inserted punctuation marks such as periods, commas, question marks, and exclamation points.

400 - Bad Request. The request failed because of a user input error. For example, the request passed audio that does not match the indicated format or failed to specify a required audio format; specified a custom language or custom acoustic model that is not in the `available` state; or experienced an inactivity timeout. Specific messages include

`Model {model} not found`
`Requested model is not available`
`This 8000hz audio input requires a narrow band model. See /v1/models for a listp of available models.`
`speaker_labels is not a supported feature for model {model}`
`keywords_threshold value must be between zero and one (inclusive)`
`word_alternatives_threshold value must be between zero and one (inclusive)`
`You cannot specify both 'customization_id' and 'language_customization_id' parameter!`
`No speech detected for 30s`
`Unable to transcode data stream application/octet-stream -> audio/l16`
`Stream was {number} bytes but needs to be at least 100 bytes.`
`keyword {keyword} length exceeds the maximum length 1024`
`low_latency is not a supported feature for model {model}`
`Character insertion bias must be a value between -1 and 1.`

Register a callback

POST /v1/register_callback

Registers a callback URL with the service for use with subsequent asynchronous recognition requests. The service attempts to register, or allowlist, the callback URL if it is not already registered by sending a `GET` request to the callback URL. The service passes a random alphanumeric challenge string via the `challenge_string` parameter of the request. The request includes an `Accept` header that specifies `text/plain` as the required response type.

To be registered successfully, the callback URL must respond to the `GET` request from the service. The response must send status code 200 and must include the challenge string in its body. Set the `Content-Type` response header to `text/plain`. Upon receiving this response, the service responds to the original registration request with response code 201.

The service sends only a single `GET` request to the callback URL. If the service does not receive a reply with a response code of 200 and a body that echoes the challenge string sent by the service within five seconds, it does not allowlist the URL; it instead sends status code 400 in response to the request to register a callback. If the requested callback URL is already allowlisted, the service responds to the initial registration request with response code 200.

If you specify a user secret with the request, the service uses it as a key to calculate an HMAC-SHA1 signature of the challenge string in its response to the `POST` request. It sends this signature in the `X-Callback-Signature` header of its `GET` request to the URL during registration. It also uses the secret to calculate a signature over the payload of every callback notification that uses the URL. The signature provides authentication and data integrity for HTTP communications.

After you successfully register a callback URL, you can use it with an indefinite number of recognition requests. You can register a maximum of 20 callback URLS in a one-hour span of time.

See also: Registering a callback URL.

Request

Path Parameters

The current status of the job:

`created`: The service successfully allowlisted the callback URL as a result of the call.
`already created`: The URL was already allowlisted.

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

500 - Internal Server Error. The service experienced an internal error.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

503 - Service Unavailable. The service is currently unavailable.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

Unregistering a callback URL.">

Unregister a callback

POST /v1/unregister_callback

Unregisters a callback URL that was previously allowlisted with a Register a callback request for use with the asynchronous interface. Once unregistered, the URL can no longer be used with asynchronous recognition requests.

See also: Unregistering a callback URL.

Request

Path Parameters

Query Parameters

callback_url

Required *

string

The callback URL that is to be unregistered.

Response

Warnings associated with the error.

Check a job method to retrieve results is more secure than receiving them via callback notification over HTTP because it provides confidentiality in addition to authentication and data integrity.

The method supports the same basic parameters as other HTTP and WebSocket recognition requests. It also supports the following parameters specific to the asynchronous interface:

`callback_url`
`events`
`user_token`
`results_ttl`

You can pass a maximum of 1 GB and a minimum of 100 bytes of audio with a request. The service automatically detects the endianness of the incoming audio and, for audio that includes multiple channels, downmixes the audio to one-channel mono during transcoding. The method returns only final results; to enable interim results, use the WebSocket API. (With the `curl` command, use the `--data-binary` option to upload the file for the request.)

See also: Creating a job.

Streaming mode

See also:

Audio formats (content types)

The service accepts audio in the following formats (MIME types).

For formats that are labeled Required, you must use the `Content-Type` header with the request to specify the format of the audio.
For all other formats, you can omit the `Content-Type` header or specify `application/octet-stream` with the header to have the service automatically detect the format of the audio. (With the `curl` command, you can specify either `"Content-Type:"` or `"Content-Type: application/octet-stream"`.)

Where indicated, the format that you specify must include the sampling rate and can optionally include the number of channels and the endianness of the audio.

`audio/alaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/basic` (Required. Use only with narrowband models.)
`audio/flac`
`audio/g729` (Use only with narrowband models.)
`audio/l16` (Required. Specify the sampling rate (`rate`) and optionally the number of channels (`channels`) and endianness (`endianness`) of the audio.)
`audio/mp3`
`audio/mpeg`
`audio/mulaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/ogg` (The service automatically detects the codec of the input audio.)
`audio/ogg;codecs=opus`
`audio/ogg;codecs=vorbis`
`audio/wav` (Provide audio with a maximum of nine channels.)
`audio/webm` (The service automatically detects the codec of the input audio.)
`audio/webm;codecs=opus`
`audio/webm;codecs=vorbis`

See also: Supported audio formats.

Large speech models and Next-generation models

`acoustic_customization_id`
`keywords` and `keywords_threshold`
`processing_metrics` and `processing_metrics_interval`
`word_alternatives_threshold`

See also:

Create a job

POST /v1/recognitions

Creates a job for a new asynchronous recognition request. The job is owned by the instance of the service whose credentials are used to create it. How you learn the status and results of a job depends on the parameters you include with the job creation request:

By callback notification: Include the `callback_url` parameter to specify a URL to which the service is to send callback notifications when the status of the job changes. Optionally, you can also include the `events` and `user_token` parameters to subscribe to specific events and to specify a string that is to be included with each notification for the job.
By polling the service: Omit the `callback_url`, `events`, and `user_token` parameters. You must then use the Check jobs or Check a job methods to check the status of the job, using the latter to retrieve the results when the job is complete.

The two approaches are not mutually exclusive. You can poll the service for job status or obtain results from the service manually even if you include a callback URL. In both cases, you can include the `results_ttl` parameter to specify how long the results are to remain available after the job is complete. Using the HTTPS Check a job method to retrieve results is more secure than receiving them via callback notification over HTTP because it provides confidentiality in addition to authentication and data integrity.

The method supports the same basic parameters as other HTTP and WebSocket recognition requests. It also supports the following parameters specific to the asynchronous interface:

`callback_url`
`events`
`user_token`
`results_ttl`

See also: Creating a job.

Streaming mode

See also:

Audio formats (content types)

The service accepts audio in the following formats (MIME types).

For formats that are labeled Required, you must use the `Content-Type` header with the request to specify the format of the audio.
For all other formats, you can omit the `Content-Type` header or specify `application/octet-stream` with the header to have the service automatically detect the format of the audio. (With the `curl` command, you can specify either `"Content-Type:"` or `"Content-Type: application/octet-stream"`.)

Where indicated, the format that you specify must include the sampling rate and can optionally include the number of channels and the endianness of the audio.

`audio/alaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/basic` (Required. Use only with narrowband models.)
`audio/flac`
`audio/g729` (Use only with narrowband models.)
`audio/l16` (Required. Specify the sampling rate (`rate`) and optionally the number of channels (`channels`) and endianness (`endianness`) of the audio.)
`audio/mp3`
`audio/mpeg`
`audio/mulaw` (Required. Specify the sampling rate (`rate`) of the audio.)
`audio/ogg` (The service automatically detects the codec of the input audio.)
`audio/ogg;codecs=opus`
`audio/ogg;codecs=vorbis`
`audio/wav` (Provide audio with a maximum of nine channels.)
`audio/webm` (The service automatically detects the codec of the input audio.)
`audio/webm;codecs=opus`
`audio/webm;codecs=vorbis`

See also: Supported audio formats.

Large speech models and Next-generation models

`acoustic_customization_id`
`keywords` and `keywords_threshold`
`processing_metrics` and `processing_metrics_interval`
`word_alternatives_threshold`

See also:

Request

Request Body

Path Parameters

Query Parameters

model

string

The model to use for speech recognition. If you omit the `model` parameter, the service uses the US English `en-US_BroadbandModel` by default.

For IBM Cloud Pak for Data, if you do not install the `en-US_BroadbandModel`, you must either specify a model with the request or specify a new default model for your installation of the service.

See also:

callback_url

string

A URL to which callback notifications are to be sent. The URL must already be successfully allowlisted by using the Register a callback method. You can include the same callback URL with any number of job creation requests. Omit the parameter to poll the service for job completion and results.

Use the `user_token` parameter to specify a unique user-specified string with each job to differentiate the callback notifications for the jobs.

events

string

If the job includes a callback URL, a comma-separated list of notification events to which to subscribe. Valid events are

`recognitions.started` generates a callback notification when the service begins to process the job.
`recognitions.completed` generates a callback notification when the job is complete. You must use the Check a job method to retrieve the results before they time out or are deleted.
`recognitions.completed_with_results` generates a callback notification when the job is complete. The notification includes the results of the request.
`recognitions.failed` generates a callback notification if the service experiences an error while processing the job.

The `recognitions.completed` and `recognitions.completed_with_results` events are incompatible. You can specify only of the two events.

If the job includes a callback URL, omit the parameter to subscribe to the default events: `recognitions.started`, `recognitions.completed`, and `recognitions.failed`. If the job does not include a callback URL, omit the parameter.

user_token

string

integer

word_alternatives_threshold

number

word_confidence

boolean

If `true`, the service returns a confidence measure in the range of 0.0 to 1.0 for each word. By default, the service returns no word confidence scores. See Word confidence.

timestamps

Note: The parameter can be used with US English, Japanese, and Korean transcription only.

See Numeric redaction.

processing_metrics

boolean

integer

Detects speech boundaries within the audio stream with better performance, improved noise suppression, faster responsiveness, and increased accuracy.

Specify `sad_module: 2`

See Speech Activity Detection (SAD).

background_audio_suppression

number

Specify a value in the range of 0.0 to 1.0:

0.0 (the default) provides no suppression (background audio suppression is disabled).
0.5 provides a reasonable level of audio suppression for general usage.
1.0 suppresses all audio (no audio is transcribed).

The values increase on a monotonic curve. Specifying one or two decimal places of precision (for example, `0.55`) is typically more than sufficient.

The parameter is supported with all large speech models, next-generation models and with most previous-generation models. See Background audio suppression and Language model support.

low_latency

boolean

The parameter is not available for large speech models and previous-generation `Broadband` and `Narrowband` models. It is available for most next-generation models.

For a list of next-generation models that support low latency, see Supported next-generation language models.
For more information about the `low_latency` parameter, see Low latency.

character_insertion_bias

number

The default bias is 0.0. The allowable range of values is -1.0 to 1.0.

Negative values bias the service to favor hypotheses with shorter strings of characters.
Positive values bias the service to favor hypotheses with longer strings of characters.

The parameter is not available for previous-generation models.

See Character insertion bias.

Response

201 - Created. The job was successfully created.

id

string

The ID of the asynchronous job.

status

string

The current status of the job:

`waiting`: The service is preparing the job for processing. The service returns this status when the job is initially created or when it is waiting for capacity to process the job. The job remains in this state until the service has the capacity to begin processing it.
`processing`: The service is actively processing the job.
`completed`: The service has finished processing the job. If the job specified a callback URL and the event `recognitions.completed_with_results`, the service sent the results with the callback notification. Otherwise, you must retrieve the results by checking the individual job.
`failed`: The job failed.

400 - Bad Request. The request failed because of a user input error. For example, the request passed audio that does not match the indicated format or failed to specify a required audio format; specified a custom language or custom acoustic model that is not in the `available` state; or specified both the `recognitions.completed` and `recognitions.completed_with_results` events. Specific messages include

`Model {model} not found`
`Requested model is not available`
`This 8000hz audio input requires a narrow band model. See /v1/models for a list of available models.`
`speaker_labels is not a supported feature for model {model}`
`keywords_threshold value must be between zero and one (inclusive)`
`word_alternatives_threshold value must be between zero and one (inclusive)`
`You cannot specify both 'customization_id' and 'language_customization_id' parameter!`
`No speech detected for 30s`
`Unable to transcode data stream application/octet-stream -> audio/l16`
`Stream was {number} bytes but needs to be at least 100 bytes.`
`keyword {keyword} length exceeds the maximum length 1024`
`low_latency is not a supported feature for model {model}`
`Character insertion bias must be a value between -1 and 1.`

Check jobs

GET /v1/recognitions

Returns the ID and status of the latest 100 outstanding jobs associated with the credentials with which it is called. The method also returns the creation and update times of each job, and, if a job was created with a callback URL and a user token, the user token for the job. To obtain the results for a job whose status is `completed` or not one of the latest 100 outstanding jobs, use the [Check a job[(#checkjob) method. A job and its results remain available until you delete them with the Delete a job method or until the job's time to live expires, whichever comes first.

Request

Response

200 - OK. The request succeeded.

recognitions

array

An array of `RecognitionJob` objects that provides the status for each of the user's current jobs. The array is empty if the user has no current jobs.

500 - Internal Server Error. The service experienced an internal error.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

503 - Service Unavailable. The service is currently unavailable.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

Check jobs method to request information about the most recent jobs associated with the calling credentials.

Check a job

GET /v1/recognitions/{id}

Returns information about the specified job. The response always includes the status of the job and its creation and update times. If the status is `completed`, the response includes the results of the recognition request. You must use credentials for the instance of the service that owns a job to list information about it.

You can use the method to retrieve the results of any job, regardless of whether it was submitted with a callback URL and the `recognitions.completed_with_results` event, and you can retrieve the results multiple times for as long as they remain available. Use the Check jobs method to request information about the most recent jobs associated with the calling credentials.

Request

Path Parameters

id

Required *

string

The identifier of the asynchronous job that is to be used for the request. You must make the request with credentials for the instance of the service that owns the job.

Query Parameters

Response

200 - OK. The request succeeded.

id

string

The ID of the asynchronous job.

status

string

The current status of the job:

`waiting`: The service is preparing the job for processing. The service returns this status when the job is initially created or when it is waiting for capacity to process the job. The job remains in this state until the service has the capacity to begin processing it.
`processing`: The service is actively processing the job.
`completed`: The service has finished processing the job. If the job specified a callback URL and the event `recognitions.completed_with_results`, the service sent the results with the callback notification. Otherwise, you must retrieve the results by checking the individual job.
`failed`: The job failed.

created

Delete a job

DELETE /v1/recognitions/{id}

Deletes the specified job. You cannot delete a job that the service is actively processing. Once you delete a job, its results are no longer available. The service automatically deletes a job and its results when the time to live for the results expires. You must use credentials for the instance of the service that owns a job to delete it.

See also: Deleting a job.

Request

Path Parameters

id

Required *

string

The identifier of the asynchronous job that is to be used for the request. You must make the request with credentials for the instance of the service that owns the job.

Query Parameters

Response

204 - No Content. The job was successfully deleted.

400 - Bad Request. The service cannot delete a job that it is actively processing:

`Unable to delete the processing job`

See also:

Large speech models and Next-generation models

`acoustic_customization_id`
`keywords` and `keywords_threshold`
`processing_metrics` and `processing_metrics_interval`
`word_alternatives_threshold`

See also:

Create a custom language model

POST /v1/customizations

Creates a new custom language model for a specified base model. The custom language model can be used only with the base model for which it is created. The model is owned by the instance of the service whose credentials are used to create it.

You can create a maximum of 1024 custom language models per owning credentials. The service returns an error if you attempt to create more than 1024 models. You do not lose any models, but you cannot create any more until your model count is below the limit.

See also:

Large speech models and Next-generation models

`acoustic_customization_id`
`keywords` and `keywords_threshold`
`processing_metrics` and `processing_metrics_interval`
`word_alternatives_threshold`

See also:

Request

The dialect of the specified language that is to be used with the custom language model. For all languages, it is always safe to omit this field. The service automatically uses the language identifier from the name of the base model. For example, the service automatically uses `en-US` for all US English models.

If you specify the `dialect` for a new custom model, follow these guidelines. For non-Spanish previous-generation models and for next-generation models, you must specify a value that matches the five-character language identifier from the name of the base model. For Spanish previous-generation models, you must specify one of the following values:

`es-ES` for Castilian Spanish (`es-ES` models)
`es-LA` for Latin American Spanish (`es-AR`, `es-CL`, `es-CO`, and `es-PE` models)
`es-US` for Mexican (North American) Spanish (`es-MX` models)

All values that you pass for the `dialect` field are case-insensitive.

description

string

A recommended description of the new custom language model. Use a localized description that matches the language of the custom model. Include a maximum of 128 characters in the description.

Response

201 - Created. The custom language model was successfully created.

customization_id

string

The customization ID (GUID) of the custom language model. The Create a custom language model method returns only this field of the object; it does not return the other fields.

created

string

The date and time in Coordinated Universal Time (UTC) at which the custom language model was created. The value is provided in full ISO 8601 format (`YYYY-MM-DDThh:mm:ss.sTZD`).

updated

string

The date and time in Coordinated Universal Time (UTC) at which the custom language model was last modified. The `created` and `updated` fields are equal when a language model is first added but has yet to be updated. The value is provided in full ISO 8601 format (YYYY-MM-DDThh:mm:ss.sTZD).

language

string

The language identifier of the custom language model (for example, `en-US`). The value matches the five-character language identifier from the name of the base model for the custom model. This value might be different from the value of the `dialect` field.

dialect

string

The dialect of the language for the custom language model. For custom models that are based on non-Spanish previous-generation models and on next-generation models, the field matches the language of the base model; for example, `en-US` for one of the US English models. For custom models that are based on Spanish previous-generation models, the field indicates the dialect with which the model was created. The value can match the name of the base model or, if it was specified by the user, can be one of the following:

`es-ES` for Castilian Spanish (`es-ES` models)
`es-LA` for Latin American Spanish (`es-AR`, `es-CL`, `es-CO`, and `es-PE` models)
`es-US` for Mexican (North American) Spanish (`es-MX` models)

The current status of the custom language model:

`pending`: The model was created but is waiting either for valid training data to be added or for the service to finish analyzing added data.
`ready`: The model contains valid data and is ready to be trained. If the model contains a mix of valid and invalid resources, you need to set the `strict` parameter to `false` for the training to proceed.
`training`: The model is currently being trained.
`available`: The model is trained and ready to use.
`upgrading`: The model is currently being upgraded.
`failed`: Training of the model failed.

progress

integer

A percentage that indicates the progress of the custom language model's current training. A value of `100` means that the model is fully trained. Note: The `progress` field does not currently reflect the progress of the training. The field changes from `0` to `100` when training is complete.

error

string

If an error occurred while adding a grammar file to the custom language model, a message that describes an `Internal Server Error` and includes the string `Cannot compile grammar`. The status of the custom model is not affected by the error, but the grammar cannot be used with the model.

warnings

string

If the request included unknown parameters, the following message: `Unexpected query parameter(s) ['parameters'] detected`, where `parameters` is a list that includes a quoted string for each unknown parameter.

400 - Bad Request. A required parameter is null or invalid. Specific failure messages include:

`Required parameter '{name}' is missing`
`Required parameter '{name}' cannot be empty string`
`Required parameter '{name}' cannot be null`
`The base model '{model_name}' is not recognized`
`Language customization is not supported for base model '{model_name}'`
`Invalid dialect value '{dialect}' specified for language '{language}'`
`You exceeded the maximum '{model_number}' of allowed custom language models. You have '{model_number}' custom language models. Please remove the models you do not need or contact the IBM speech support team to apply for an exception.`

Language support for customization">

List custom language models

GET /v1/customizations

Lists information about all custom language models that are owned by an instance of the service. Use the `language` parameter to see all custom language models for the specified language. Omit the parameter to see all custom language models for all languages. You must use credentials for the instance of the service that owns a model to list information about it.

See also:

Request

Path Parameters

Query Parameters

language

string

The identifier of the language for which custom language or custom acoustic models are to be returned. Specify the five-character language identifier; for example, specify `en-US` to see all custom language or custom acoustic models that are based on US English models. Omit the parameter to see all custom language or custom acoustic models that are owned by the requesting credentials.

To determine the languages for which customization is available, see Language support for customization.

Response

200 - OK. The request succeeded.

customizations

array

An array of `LanguageModel` objects that provides information about each available custom language model. The array is empty if the requesting credentials own no custom language models (if no language is specified) or own no custom language models for the specified language.

400 - Bad Request. A required parameter is null or invalid. Specific failure messages include:

`Language '{language}' is not supported for customization`.

Language support for customization">

Get a custom language model

GET /v1/customizations/{customization_id}

Gets information about a specified custom language model. You must use credentials for the instance of the service that owns a model to list information about it.

See also:

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The request succeeded.

customization_id

string

The customization ID (GUID) of the custom language model. The Create a custom language model method returns only this field of the object; it does not return the other fields.

created

string

string

The GUID of the credentials for the instance of the service that owns the custom language model.

name

string

The name of the custom language model.

description

string

The description of the custom language model.

base_model_name

string

The name of the language model for which the custom language model was created.

status

string

The current status of the custom language model:

`pending`: The model was created but is waiting either for valid training data to be added or for the service to finish analyzing added data.
`ready`: The model contains valid data and is ready to be trained. If the model contains a mix of valid and invalid resources, you need to set the `strict` parameter to `false` for the training to proceed.
`training`: The model is currently being trained.
`available`: The model is trained and ready to use.
`upgrading`: The model is currently being upgraded.
`failed`: Training of the model failed.

progress

integer

error

string

warnings

string

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

Language support for customization">

Delete a custom language model

DELETE /v1/customizations/{customization_id}

Deletes an existing custom language model. The custom model cannot be deleted if another request, such as adding a corpus or grammar to the model, is currently being processed. You must use credentials for the instance of the service that owns a model to delete it.

See also:

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The custom language model was successfully deleted.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials, including the case where the custom model does not exist:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

The method returns a `LanguageModel` object that includes `status` and `progress` fields. A status of `available` means that the custom model is trained and ready to use. The service cannot accept subsequent training requests or requests to add new resources until the existing request completes.

For custom models that are based on improved base language models, training also performs an automatic upgrade to a newer version of the base model. You do not need to use the Upgrade a custom language model method to perform the upgrade.

See also:

Training failures

Training can fail to start for the following reasons:

The service is currently handling another request for the custom model, such as another training request or a request to add a corpus or grammar to the model.
No training data have been added to the custom model.
The custom model contains one or more invalid corpora, grammars, or words (for example, a custom word has an invalid sounds-like pronunciation). You can correct the invalid resources or set the `strict` parameter to `false` to exclude the invalid resources from the training. The model must contain at least one valid resource for training to succeed.">

Train a custom language model

POST /v1/customizations/{customization_id}/train

Initiates the training of a custom language model with new resources such as corpora, grammars, and custom words. After adding, modifying, or deleting resources for a custom language model, use this method to begin the actual training of the model on the latest data. You can specify whether the custom language model is to be trained with all words from its words resource or only with words that were added or modified by the user directly. You must use credentials for the instance of the service that owns a model to train it.

The training method is asynchronous. It can take on the order of minutes to complete depending on the amount of data on which the service is being trained and the current load on the service. The method returns an HTTP 200 response code to indicate that the training process has begun.

You can monitor the status of the training by using the Get a custom language model method to poll the model's status. Use a loop to check the status every 10 seconds. If you added custom words directly to a custom model that is based on a next-generation model, allow for some minutes of extra training time for the model.

See also:

Training failures

Training can fail to start for the following reasons:

The service is currently handling another request for the custom model, such as another training request or a request to add a corpus or grammar to the model.
No training data have been added to the custom model.
The custom model contains one or more invalid corpora, grammars, or words (for example, a custom word has an invalid sounds-like pronunciation). You can correct the invalid resources or set the `strict` parameter to `false` to exclude the invalid resources from the training. The model must contain at least one valid resource for training to succeed.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

word_type_to_add

string

For custom models that are based on previous-generation models, the type of words from the custom language model's words resource on which to train the model:

`all` (the default) trains the model on all new words, regardless of whether they were extracted from corpora or grammars or were added or modified by the user.
`user` trains the model only on custom words that were added or modified by the user directly. The model is not trained on new words extracted from corpora or grammars.

For custom models that are based on large speech models and next-generation models, the service ignores the `word_type_to_add` parameter. The words resource contains only custom words that the user adds or modifies directly, so the parameter is unnecessary.

customization_weight

number

Specifies a customization weight for the custom language model. The customization weight tells the service how much weight to give to words from the custom language model compared to those from the base model for speech recognition. Specify a value between 0.0 and 1.0. The default value is:

0.5 for large speech models
0.3 for previous-generation models
0.2 for most next-generation models
0.1 for next-generation English and Japanese models

The default value yields the best performance in general. Assign a higher value if your audio makes frequent use of OOV words from the custom model. Use caution when setting the weight: a higher value can improve the accuracy of phrases from the custom model's domain, but it can negatively affect performance on non-domain phrases.

The value that you assign is used for all recognition requests that use the model. You can override it for any recognition request by specifying a customization weight for that request.

See Using customization weight.

strict

boolean

If `false`, allows training of the custom language model to proceed as long as the model contains at least one valid resource. The method returns an array of `TrainingWarning` objects that lists any invalid resources. By default (`true`), training of a custom language model fails (status code 400) if the model contains one or more invalid resources (corpus files, grammar files, or custom words).

force

boolean

If `true`, forces the training of the custom language model regardless of whether it contains any changes (is in the `ready` or `available` state). By default (`false`), the model must be in the `ready` state to be trained. You can use the parameter to train and thus upgrade a custom model that is based on an improved next-generation model. The parameter is available only for IBM Cloud, not for IBM Cloud Pak for Data.

See Upgrading a custom language model based on an improved next-generation model.

Response

200 - OK. Training of the custom language model started successfully.

warnings

array

An array of `TrainingWarning` objects that lists any invalid resources contained in the custom model. For custom language models, invalid resources are grouped and identified by type of resource. The method can return warnings only if the `strict` parameter is set to `false`.

400 - Bad Request. A required parameter is null or invalid, the custom model is not ready to be trained, or the total number of words or OOV words exceeds the maximum threshold. Specific failure messages include:

`No input data available for running training`
`Fix errors in the following words: [{words}] before training`
`Total number of words {number} exceeds maximum allowed`
`Total number of OOV words {number} exceeds {maximum}`
`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Language support for customization">

Reset a custom language model

POST /v1/customizations/{customization_id}/reset

Resets a custom language model by removing all corpora, grammars, and words from the model. Resetting a custom language model initializes the model to its state when it was first created. Metadata such as the name and language of the model are preserved, but the model's words resource is removed and must be re-created. You must use credentials for the instance of the service that owns a model to reset it.

See also:

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The custom language model was successfully reset.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

While it is being upgraded, the custom model has the status `upgrading`. When the upgrade is complete, the model resumes the status that it had prior to upgrade. The service cannot accept subsequent requests for the model until the upgrade completes.

For custom models that are based on improved base language models, the Train a custom language model method also performs an automatic upgrade to a newer version of the base model. You do not need to use the upgrade method.

See also:

Upgrade a custom language model

POST /v1/customizations/{customization_id}/upgrade_model

Initiates the upgrade of a custom language model to the latest version of its base language model. The upgrade method is asynchronous. It can take on the order of minutes to complete depending on the amount of data in the custom model and the current load on the service. A custom model must be in the `ready` or `available` state to be upgraded. You must use credentials for the instance of the service that owns a model to upgrade it.

The method returns an HTTP 200 response code to indicate that the upgrade process has begun successfully. You can monitor the status of the upgrade by using the Get a custom language model method to poll the model's status. The method returns a `LanguageModel` object that includes `status` and `progress` fields. Use a loop to check the status every 10 seconds.

See also:

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. Upgrade of the custom language model started successfully.

400 - Bad Request. The specified customization ID is invalid or the specified custom model cannot be upgraded:

`Malformed GUID: '{customization_id}'`
`Custom model is up-to-date`
`No input data available to upgrade the model`
`Cannot upgrade failed custom model`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

List corpora

GET /v1/customizations/{customization_id}/corpora

Lists information about all corpora from a custom language model. The information includes the name, status, and total number of words for each corpus. For custom models that are based on previous-generation models, it also includes the number of out-of-vocabulary (OOV) words from the corpus. You must use credentials for the instance of the service that owns a model to list its corpora.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The request succeeded.

corpora

array

An array of `Corpus` objects that provides information about the corpora for the custom model. The array is empty if the custom model has no corpora.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

For custom models that are based on large speech models, the service parses and extracts word sequences from one or multiple corpora files. The characters help the service learn and predict character sequences from audio.

For custom models that are based on previous-generation models, the service auto-populates the model's words resource with words from the corpus that are not found in its base vocabulary. These words are referred to as out-of-vocabulary (OOV) words. After adding a corpus, you must validate the words resource to ensure that each OOV word's definition is complete and valid. You can use the List custom words method to examine the words resource. You can use other words method to eliminate typos and modify how words are pronounced and displayed as needed.

To add a corpus file that has the same name as an existing corpus, set the `allow_overwrite` parameter to `true`; otherwise, the request fails. Overwriting an existing corpus causes the service to process the corpus text file and extract its data anew. For a custom model that is based on a previous-generation model, the service first removes any OOV words that are associated with the existing corpus from the model's words resource unless they were also added by another corpus or grammar, or they have been modified in some way with the Add custom words or Add a custom word method.

The service limits the overall amount of data that you can add to a custom model to a maximum of 10 million total words from all sources combined. For a custom model that is based on a previous-generation model, you can add no more than 90 thousand custom (OOV) words to a model. This includes words that the service extracts from corpora and grammars, and words that you add directly.

See also:

Add a corpus

POST /v1/customizations/{customization_id}/corpora/{corpus_name}

Adds a single corpus text file of new training data to a custom language model. Use multiple requests to submit multiple corpus text files. You must use credentials for the instance of the service that owns a model to add a corpus to it. Adding a corpus does not affect the custom language model until you train the model for the new data by using the Train a custom language model method.

Submit a plain text file that contains sample sentences from the domain of interest to enable the service to parse the words in context. The more sentences you add that represent the context in which speakers use words from the domain, the better the service's recognition accuracy.

The call returns an HTTP 201 response code if the corpus is valid. The service then asynchronously processes and automatically extracts data from the contents of the corpus. This operation can take on the order of minutes to complete depending on the current load on the service, the total number of words in the corpus, and, for custom models that are based on previous-generation models, the number of new (out-of-vocabulary) words in the corpus. You cannot submit requests to add additional resources to the custom model or to train the model until the service's analysis of the corpus for the current request completes. Use the Get a corpus method to check the status of the analysis.

See also:

Request

Request Body

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

corpus_name

Required *

string

The name of the new corpus for the custom language model. Use a localized name that matches the language of the custom model and reflects the contents of the corpus.

Include a maximum of 128 characters in the name.
Do not use characters that need to be URL-encoded. For example, do not use spaces, slashes, backslashes, colons, ampersands, double quotes, plus signs, equals signs, questions marks, and so on in the name. (The service does not prevent the use of these characters. But because they must be URL-encoded wherever used, their use is strongly discouraged.)
Do not use the name of an existing corpus or grammar that is already defined for the custom model.
Do not use the name `user`, which is reserved by the service to denote custom words that are added or modified by the user.
Do not use the name `base_lm` or `default_lm`. Both names are reserved for future use by the service.

Query Parameters

allow_overwrite

boolean

If `true`, the specified corpus overwrites an existing corpus with the same name. If `false`, the request fails if a corpus with the same name already exists. The parameter has no effect if a corpus with the same name does not already exist.

Response

201 - Created. Addition of the corpus data was successfully started. The service is analyzing the data.

400 - Bad Request. A required parameter is null or invalid, the specified corpus name already exists, or the custom model needs to be upgraded, among other possibilities. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Corpus file not specified or empty`
`Corpus '{corpus_name}' already exists - change its name, remove existing file before adding new one, or overwrite existing file by setting 'allow_overwrite' to 'true'`
`Grammar exists with corpus name '{corpus_name}'. Please use different name.`
`TOTAL_NUMBER_OF_OOV_WORDS_EXCEEDS_MAXIMUM_ALLOWED_FORMAT: "Total number of OOV words {total_number} exceeds {maximum_allowed}"`
`Analysis of corpus '{corpus_name}' failed due to {error_message}. Please fix the error then add the corpus again by setting the 'allow_overwrite' flag to 'true'.`, where `{error_message}` is a message of the form `{"code": 404, "error": "Model en-US_BroadbandModel (version: en-US_BroadbandModel.{version}) not found", "code_description": "Not Found"}`. Upgrade the custom language model to the latest version of its base language model, and then add the corpus to the custom model.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. The corpus name includes characters that need to be URL-encoded.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Get a corpus

GET /v1/customizations/{customization_id}/corpora/{corpus_name}

Gets information about a corpus from a custom language model. The information includes the name, status, and total number of words for the corpus. For custom models that are based on previous-generation models, it also includes the number of out-of-vocabulary (OOV) words from the corpus. You must use credentials for the instance of the service that owns a model to list its corpora.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

corpus_name

Required *

string

The name of the corpus for the custom language model.

Query Parameters

Response

200 - OK. The request succeeded.

name

string

The name of the corpus.

total_words

integer

The total number of words in the corpus. The value is `0` while the corpus is being processed.

out_of_vocabulary_words

integer

For custom models that are based on large speech models and previous-generation models, the number of OOV words extracted from the corpus. The value is `0` while the corpus is being processed.

For custom models that are based on next-generation models, no OOV words are extracted from corpora, so the value is always `0`.

status

string

The status of the corpus:

`analyzed`: The service successfully analyzed the corpus. The custom model can be trained with data from the corpus.
`being_processed`: The service is still analyzing the corpus. The service cannot accept requests to add new resources or to train the custom model.
`undetermined`: The service encountered an error while processing the corpus. The `error` field describes the failure.

error

string

If the status of the corpus is `undetermined`, the following message: `Analysis of corpus 'name' failed. Please try adding the corpus again by setting the 'allow_overwrite' flag to 'true'`.

400 - Bad Request. The specified customization ID or corpus name is invalid, including the case where the corpus does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for corpus name '{corpus_name}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

Delete a corpus

DELETE /v1/customizations/{customization_id}/corpora/{corpus_name}

Deletes an existing corpus from a custom language model. Removing a corpus does not affect the custom model until you train the model with the Train a custom language model method. You must use credentials for the instance of the service that owns a model to delete its corpora.

For custom models that are based on previous-generation models, the service removes any out-of-vocabulary (OOV) words that are associated with the corpus from the custom model's words resource unless they were also added by another corpus or grammar, or they were modified in some way with the Add custom words or Add a custom word method.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

corpus_name

Required *

string

The name of the corpus for the custom language model.

Query Parameters

Response

200 - OK. The corpus was successfully deleted from the custom language model.

400 - Bad Request. The specified customization ID or corpus name is invalid, including the case where the corpus does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for corpus name '{corpus_name}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. No corpus name was specified with the request.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

List custom words

GET /v1/customizations/{customization_id}/words

Lists information about custom words from a custom language model. You can list all words from the custom model's words resource, only custom words that were added or modified by the user, or, for a custom model that is based on a previous-generation model, only out-of-vocabulary (OOV) words that were extracted from corpora or are recognized by grammars. For a custom model that is based on a next-generation model, you can list all words or only those words that were added directly by a user, which return the same results.

You can also indicate the order in which the service is to return words; by default, the service lists words in ascending alphabetical order. You must use credentials for the instance of the service that owns a model to list information about its words.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

word_type

string

The type of words to be listed from the custom language model's words resource:

`all` (the default) shows all words.
`user` shows only custom words that were added or modified by the user directly.
`corpora` shows only OOV that were extracted from corpora.
`grammars` shows only OOV words that are recognized by grammars.

For a custom model that is based on a next-generation model, only `all` and `user` apply. Both options return the same results. Words from other sources are not added to custom models that are based on next-generation models.

sort

string

Indicates the order in which the words are to be listed, `alphabetical` or by `count`. You can prepend an optional `+` or `-` to an argument to indicate whether the results are to be sorted in ascending or descending order. By default, words are sorted in ascending alphabetical order. For alphabetical ordering, the lexicographical precedence is numeric values, uppercase letters, and lowercase letters. For count ordering, values with the same count are ordered alphabetically. With the `curl` command, URL-encode the `+` symbol as `%2B`.

Response

200 - OK. The request succeeded.

words

array

An array of `Word` objects that provides information about each word in the custom model's words resource. The array is empty if the custom model has no words.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

You add custom words by providing a `CustomWords` object, which is an array of `CustomWord` objects, one per word. Use the object's `word` parameter to identify the word that is to be added. You can also provide one or both of the optional `display_as` or `sounds_like` fields for each word.

The `display_as` field provides a different way of spelling the word in a transcript. Use the parameter when you want the word to appear different from its usual representation or from its spelling in training data. For example, you might indicate that the word `IBM` is to be displayed as `IBM™`.
The `sounds_like` field provides an array of one or more pronunciations for the word. Use the parameter to specify how the word can be pronounced by users. Use the parameter for words that are difficult to pronounce, foreign words, acronyms, and so on. For example, you might specify that the word `IEEE` can sound like `I triple E`. You can specify a maximum of five sounds-like pronunciations for a word. For a custom model that is based on a previous-generation model, if you omit the `sounds_like` field, the service attempts to set the field to its pronunciation of the word. It cannot generate a pronunciation for all words, so you must review the word's definition to ensure that it is complete and valid.
The `mapping_only` field provides parameter for custom words. You can use the 'mapping_only' key in custom words as a form of post processing. This key parameter has a boolean value to determine whether 'sounds_like' (for non-Japanese models) or word (for Japanese) is not used for the model fine-tuning, but for the replacement for 'display_as'. This feature helps you when you use custom words exclusively to map 'sounds_like' (or word) to 'display_as' value. When you use custom words solely for post-processing purposes that does not need fine-tuning.

If you add a custom word that already exists in the words resource for the custom model, the new definition overwrites the existing data for the word. If the service encounters an error with the input data, it returns a failure code and does not add any of the words to the words resource.

The call returns an HTTP 201 response code if the input data is valid. It then asynchronously processes the words to add them to the model's words resource. The time that it takes for the analysis to complete depends on the number of new words that you add but is generally faster than adding a corpus or grammar.

You can monitor the status of the request by using the Get a custom language model method to poll the model's status. Use a loop to check the status every 10 seconds. The method returns a `Customization` object that includes a `status` field. A status of `ready` means that the words have been added to the custom model. The service cannot accept requests to add new data or to train the model until the existing request completes.

You can use the List custom words or Get a custom word method to review the words that you add. Words with an invalid `sounds_like` field include an `error` field that describes the problem. You can use other words-related methods to correct errors, eliminate typos, and modify how words are pronounced as needed.

See also:

Add custom words

POST /v1/customizations/{customization_id}/words

Adds one or more custom words to a custom language model. You can use this method to add words or to modify existing words in a custom model's words resource. For custom models that are based on previous-generation models, the service populates the words resource for a custom model with out-of-vocabulary (OOV) words from each corpus or grammar that is added to the model. You can use this method to modify OOV words in the model's words resource.

For a custom model that is based on a previous-generation model, the words resource for a model can contain a maximum of 90 thousand custom (OOV) words. This includes words that the service extracts from corpora and grammars and words that you add directly.

You must use credentials for the instance of the service that owns a model to add or modify custom words for the model. Adding or modifying custom words does not affect the custom model until you train the model for the new data by using the Train a custom language model method.

The `display_as` field provides a different way of spelling the word in a transcript. Use the parameter when you want the word to appear different from its usual representation or from its spelling in training data. For example, you might indicate that the word `IBM` is to be displayed as `IBM™`.
The `sounds_like` field provides an array of one or more pronunciations for the word. Use the parameter to specify how the word can be pronounced by users. Use the parameter for words that are difficult to pronounce, foreign words, acronyms, and so on. For example, you might specify that the word `IEEE` can sound like `I triple E`. You can specify a maximum of five sounds-like pronunciations for a word. For a custom model that is based on a previous-generation model, if you omit the `sounds_like` field, the service attempts to set the field to its pronunciation of the word. It cannot generate a pronunciation for all words, so you must review the word's definition to ensure that it is complete and valid.
The `mapping_only` field provides parameter for custom words. You can use the 'mapping_only' key in custom words as a form of post processing. This key parameter has a boolean value to determine whether 'sounds_like' (for non-Japanese models) or word (for Japanese) is not used for the model fine-tuning, but for the replacement for 'display_as'. This feature helps you when you use custom words exclusively to map 'sounds_like' (or word) to 'display_as' value. When you use custom words solely for post-processing purposes that does not need fine-tuning.

See also:

Request

Request Body

CustomWords

A `CustomWords` object that provides information about one or more custom words that are to be added to or updated in the custom language model.

words

Required *

array

An array of `CustomWord` objects that provides information about each custom word that is to be added to or updated in the custom language model.

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

201 - Created. Addition of the custom words was successfully started. The service is analyzing the data.

400 - Bad Request. A required parameter is null or invalid, the JSON input is invalid, or the maximum number of sounds-like pronunciations for a word is exceeded. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Required property '{property}' is missing in JSON '{JSON}'`
`Word '{word}' contains invalid character character`
`Maximum number of sounds-like for a word exceeded`
`Maximum number of allowed phones of one item of sounds_like for word '{word}' exceeded`
`Malformed JSON: '{JSON}'`
`Wrong type of parameter '{parameter}' detected in the passed JSON`
`TOTAL_NUMBER_OF_OOV_WORDS_EXCEEDS_MAXIMUM_ALLOWED_FORMAT: "Total number of OOV words {total_number} exceeds {maximum_allowed}"`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Use the `word_name` parameter to specify the custom word that is to be added or modified. Use the `CustomWord` object to provide one or both of the optional `display_as` or `sounds_like` fields for the word.

The `display_as` field provides a different way of spelling the word in a transcript. Use the parameter when you want the word to appear different from its usual representation or from its spelling in training data. For example, you might indicate that the word `IBM` is to be displayed as `IBM™`.
The `sounds_like` field provides an array of one or more pronunciations for the word. Use the parameter to specify how the word can be pronounced by users. Use the parameter for words that are difficult to pronounce, foreign words, acronyms, and so on. For example, you might specify that the word `IEEE` can sound like `i triple e`. You can specify a maximum of five sounds-like pronunciations for a word. For custom models that are based on previous-generation models, if you omit the `sounds_like` field, the service attempts to set the field to its pronunciation of the word. It cannot generate a pronunciation for all words, so you must review the word's definition to ensure that it is complete and valid.

If you add a custom word that already exists in the words resource for the custom model, the new definition overwrites the existing data for the word. If the service encounters an error, it does not add the word to the words resource. Use the Get a custom word method to review the word that you add.

See also:

Add a custom word

PUT /v1/customizations/{customization_id}/words/{word_name}

Adds a custom word to a custom language model. You can use this method to add a word or to modify an existing word in the words resource. For custom models that are based on previous-generation models, the service populates the words resource for a custom model with out-of-vocabulary (OOV) words from each corpus or grammar that is added to the model. You can use this method to modify OOV words in the model's words resource.

For a custom model that is based on a previous-generation models, the words resource for a model can contain a maximum of 90 thousand custom (OOV) words. This includes words that the service extracts from corpora and grammars and words that you add directly.

You must use credentials for the instance of the service that owns a model to add or modify a custom word for the model. Adding or modifying a custom word does not affect the custom model until you train the model for the new data by using the Train a custom language model method.

The `display_as` field provides a different way of spelling the word in a transcript. Use the parameter when you want the word to appear different from its usual representation or from its spelling in training data. For example, you might indicate that the word `IBM` is to be displayed as `IBM™`.
The `sounds_like` field provides an array of one or more pronunciations for the word. Use the parameter to specify how the word can be pronounced by users. Use the parameter for words that are difficult to pronounce, foreign words, acronyms, and so on. For example, you might specify that the word `IEEE` can sound like `i triple e`. You can specify a maximum of five sounds-like pronunciations for a word. For custom models that are based on previous-generation models, if you omit the `sounds_like` field, the service attempts to set the field to its pronunciation of the word. It cannot generate a pronunciation for all words, so you must review the word's definition to ensure that it is complete and valid.

If you add a custom word that already exists in the words resource for the custom model, the new definition overwrites the existing data for the word. If the service encounters an error, it does not add the word to the words resource. Use the Get a custom word method to review the word that you add.

See also:

Request

Request Body

CustomWord

A `CustomWord` object that provides information about the specified custom word. Specify an empty object to add a word with no sounds-like or display-as information.

word

string

For the Add custom words method, you must specify the custom word that is to be added to or updated in the custom model. Do not use characters that need to be URL-encoded, for example, spaces, slashes, backslashes, colons, ampersands, double quotes, plus signs, equals signs, or question marks. Use a `-` (dash) or `_` (underscore) to connect the tokens of compound words. A Japanese custom word can include at most 25 characters, not including leading or trailing spaces.

Omit this parameter for the Add a custom word method.

mapping_only

boolean

Parameter for custom words. You can use the 'mapping_only' key in custom words as a form of post processing. This key parameter has a boolean value to determine whether 'sounds_like' (for non-Japanese models) or word (for Japanese) is not used for the model fine-tuning, but for the replacement for 'display_as'. This feature helps you when you use custom words exclusively to map 'sounds_like' (or word) to 'display_as' value. When you use custom words solely for post-processing purposes that does not need fine-tuning.

sounds_like

array

As array of sounds-like pronunciations for the custom word. Specify how words that are difficult to pronounce, foreign words, acronyms, and so on can be pronounced by users.

For custom models that are based on previous-generation models, for a word that is not in the service's base vocabulary, omit the parameter to have the service automatically generate a sounds-like pronunciation for the word.
For a word that is in the service's base vocabulary, use the parameter to specify additional pronunciations for the word. You cannot override the default pronunciation of a word; pronunciations you add augment the pronunciation from the base vocabulary.

A word can have at most five sounds-like pronunciations. A pronunciation can include at most 40 characters, not including leading or trailing spaces. A Japanese pronunciation can include at most 25 characters, not including leading or trailing spaces.

display_as

400 - Bad Request. The specified customization ID is invalid, or the maximum number of sounds-like pronunciations for a word is exceeded. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Maximum number of sounds-like for a word exceeded`
`Maximum number of allowed phones of one item of sounds_like for word '{word}' exceeded`
`TOTAL_NUMBER_OF_OOV_WORDS_EXCEEDS_MAXIMUM_ALLOWED_FORMAT: "Total number of OOV words {total_number} exceeds {maximum_allowed}"`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. The word name includes characters that need to be URL-encoded.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Get a custom word

GET /v1/customizations/{customization_id}/words/{word_name}

Gets information about a custom word from a custom language model. You must use credentials for the instance of the service that owns a model to list information about its words.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

word_name

Required *

string

The custom word that is to be read from the custom language model. URL-encode the word if it includes non-ASCII characters. For more information, see Character encoding.

Query Parameters

Response

200 - OK. The request succeeded.

word

string

A word from the custom model's words resource. The spelling of the word is used to train the model.

mapping_only

boolean

(Optional) Parameter for custom words. You can use the 'mapping_only' key in custom words as a form of post processing. A boolean value that indicates whether the added word should be used to fine-tune the mode for selected next-gen models. This field appears in the response body only when it's 'For a custom model that is based on a previous-generation model', the mapping_only field is populated with the value set by the user, but would not be used.

sounds_like

array

An array of as many as five pronunciations for the word.

For a custom model that is based on a previous-generation model, in addition to sounds-like pronunciations that were added by a user, the array can include a sounds-like pronunciation that is automatically generated by the service if none is provided when the word is added to the custom model.
For a custom model that is based on a next-generation model, the array can include only sounds-like pronunciations that were added by a user.

display_as

string

The spelling of the word that the service uses to display the word in a transcript.

For a custom model that is based on a previous-generation model, the field can contain an empty string if no display-as value is provided for a word that exists in the service's base vocabulary. In this case, the word is displayed as it is spelled.
For a custom model that is based on a next-generation model, the service uses the spelling of the word as the value of the display-as field when the word is added to the model.

count

integer

For a custom model that is based on a previous-generation model, a sum of the number of times the word is found across all corpora and grammars. For example, if the word occurs five times in one corpus and seven times in another, its count is `12`. If you add a custom word to a model before it is added by any corpora or grammars, the count begins at `1`; if the word is added from a corpus or grammar first and later modified, the count reflects only the number of times it is found in corpora and grammars.

For a custom model that is based on a next-generation model, the `count` field for any word is always `1`.

source

array

An array of sources that describes how the word was added to the custom model's words resource.

For a custom model that is based on previous-generation model, the field includes the name of each corpus and grammar from which the service extracted the word. For OOV that are added by multiple corpora or grammars, the names of all corpora and grammars are listed. If you modified or added the word directly, the field includes the string `user`.
For a custom model that is based on a next-generation model, this field shows only `user` for custom words that were added directly to the custom model. Words from corpora and grammars are not added to the words resource for custom models that are based on next-generation models.

error

array

If the service discovered one or more problems that you need to correct for the word's definition, an array that describes each of the errors.

400 - Bad Request. The specified customization ID or word is invalid, including the case where the word does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for word '{word}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

Delete a custom word

DELETE /v1/customizations/{customization_id}/words/{word_name}

Deletes a custom word from a custom language model. You can remove any word that you added to the custom model's words resource via any means. However, if the word also exists in the service's base vocabulary, the service removes the word only from the words resource; the word remains in the base vocabulary. Removing a custom word does not affect the custom model until you train the model with the Train a custom language model method. You must use credentials for the instance of the service that owns a model to delete its words.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

word_name

Required *

string

The custom word that is to be deleted from the custom language model. URL-encode the word if it includes non-ASCII characters. For more information, see Character encoding.

Query Parameters

Response

200 - OK. The custom word was successfully deleted from the custom language model.

400 - Bad Request. The specified customization ID or word is invalid, including the case where the word does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for word '{word}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. No word name was specified with the request.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Language support for customization">

List grammars

GET /v1/customizations/{customization_id}/grammars

Lists information about all grammars from a custom language model. For each grammar, the information includes the name, status, and (for grammars that are based on previous-generation models) the total number of out-of-vocabulary (OOV) words. You must use credentials for the instance of the service that owns a model to list its grammars.

See also:

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The request succeeded.

grammars

array

An array of `Grammar` objects that provides information about the grammars for the custom model. The array is empty if the custom model has no grammars.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

For grammars that are based on previous-generation models, the service populates the model's words resource with any word that is recognized by the grammar that is not found in the model's base vocabulary. These are referred to as out-of-vocabulary (OOV) words. You can use the List custom words method to examine the words resource and use other words-related methods to eliminate typos and modify how words are pronounced as needed. For grammars that are based on next-generation models, the service extracts no OOV words from the grammars.

To add a grammar that has the same name as an existing grammar, set the `allow_overwrite` parameter to `true`; otherwise, the request fails. Overwriting an existing grammar causes the service to process the grammar file and extract OOV words anew. Before doing so, it removes any OOV words associated with the existing grammar from the model's words resource unless they were also added by another resource or they have been modified in some way with the Add custom words or Add a custom word method.

For grammars that are based on previous-generation models, the service limits the overall amount of data that you can add to a custom model to a maximum of 10 million total words from all sources combined. Also, you can add no more than 90 thousand OOV words to a model. This includes words that the service extracts from corpora and grammars and words that you add directly.

See also:

Add a grammar

POST /v1/customizations/{customization_id}/grammars/{grammar_name}

Adds a single grammar file to a custom language model. Submit a plain text file in UTF-8 format that defines the grammar. Use multiple requests to submit multiple grammar files. You must use credentials for the instance of the service that owns a model to add a grammar to it. Adding a grammar does not affect the custom language model until you train the model for the new data by using the Train a custom language model method.

The call returns an HTTP 201 response code if the grammar is valid. The service then asynchronously processes the contents of the grammar and automatically extracts new words that it finds. This operation can take a few seconds or minutes to complete depending on the size and complexity of the grammar, as well as the current load on the service. You cannot submit requests to add additional resources to the custom model or to train the model until the service's analysis of the grammar for the current request completes. Use the Get a grammar method to check the status of the analysis.

See also:

Request

Request Body

A plain text file that contains the grammar in the format specified by the `Content-Type` header. Encode the file in UTF-8 (ASCII is a subset of UTF-8). Using any other encoding can lead to issues when compiling the grammar or to unexpected results in decoding. The service ignores an encoding that is specified in the header of the grammar.

With the `curl` command, use the `--data-binary` option to upload the file for the request.

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

grammar_name

Required *

string

The name of the new grammar for the custom language model. Use a localized name that matches the language of the custom model and reflects the contents of the grammar.

Include a maximum of 128 characters in the name.
Do not use characters that need to be URL-encoded. For example, do not use spaces, slashes, backslashes, colons, ampersands, double quotes, plus signs, equals signs, questions marks, and so on in the name. (The service does not prevent the use of these characters. But because they must be URL-encoded wherever used, their use is strongly discouraged.)
Do not use the name of an existing grammar or corpus that is already defined for the custom model.
Do not use the name `user`, which is reserved by the service to denote custom words that are added or modified by the user.
Do not use the name `base_lm` or `default_lm`. Both names are reserved for future use by the service.

Query Parameters

allow_overwrite

boolean

If `true`, the specified grammar overwrites an existing grammar with the same name. If `false`, the request fails if a grammar with the same name already exists. The parameter has no effect if a grammar with the same name does not already exist.

Response

201 - Created. Addition of the grammar was successfully started. The service is analyzing the data.

400 - Bad Request. A required parameter is null or invalid, the specified grammar name already exists, or the custom model needs to be upgraded, among other possibilities. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Grammar file not specified or empty`
`Grammar '{grammar_name}' already exists - change its name, remove existing grammar before adding new one, or overwrite existing grammar by setting 'allow_overwrite' to 'true'`
`Corpus exists with grammar name '{grammar_name}'. Please use different name.`
`TOTAL_NUMBER_OF_OOV_WORDS_EXCEEDS_MAXIMUM_ALLOWED_FORMAT: "Total number of OOV words {total_number} exceeds {maximum_allowed}"`
`Analysis of grammar '{grammar_name}' failed due to {error_message}. Please fix the error then add the grammar again by setting the 'allow_overwrite' flag to 'true'.`, where `{error_message}` is a message of the form `{"code": 404, "error": "Model en-US_BroadbandModel (version: en-US_BroadbandModel.{version}) not found", "code_description": "Not Found"}`. Upgrade the custom language model to the latest version of its base language model, and then add the grammar to the custom model.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. The grammar name includes characters that need to be URL-encoded.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Language support for customization">

Get a grammar

GET /v1/customizations/{customization_id}/grammars/{grammar_name}

Gets information about a grammar from a custom language model. For each grammar, the information includes the name, status, and (for grammars that are based on previous-generation models) the total number of out-of-vocabulary (OOV) words. You must use credentials for the instance of the service that owns a model to list its grammars.

See also:

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

grammar_name

Required *

string

The name of the grammar for the custom language model.

Query Parameters

Response

200 - OK. The request succeeded.

name

string

The name of the grammar.

out_of_vocabulary_words

integer

For custom models that are based on previous-generation models, the number of OOV words extracted from the grammar. The value is `0` while the grammar is being processed.

For custom models that are based on next-generation models, no OOV words are extracted from grammars, so the value is always `0`.

status

string

The status of the grammar:

`analyzed`: The service successfully analyzed the grammar. The custom model can be trained with data from the grammar.
`being_processed`: The service is still analyzing the grammar. The service cannot accept requests to add new resources or to train the custom model.
`undetermined`: The service encountered an error while processing the grammar. The `error` field describes the failure.

error

string

If the status of the grammar is `undetermined`, the following message: `Analysis of grammar '{grammar_name}' failed. Please try fixing the error or adding the grammar again by setting the 'allow_overwrite' flag to 'true'.`

400 - Bad Request. The specified customization ID or grammar name is invalid, including the case where the grammar does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for grammar name '{grammar_name}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

Language support for customization">

Delete a grammar

DELETE /v1/customizations/{customization_id}/grammars/{grammar_name}

Deletes an existing grammar from a custom language model. For grammars that are based on previous-generation models, the service removes any out-of-vocabulary (OOV) words associated with the grammar from the custom model's words resource unless they were also added by another resource or they were modified in some way with the Add custom words or Add a custom word method. Removing a grammar does not affect the custom model until you train the model with the Train a custom language model method. You must use credentials for the instance of the service that owns a model to delete its grammar.

See also:

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom language model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

grammar_name

Required *

string

The name of the grammar for the custom language model.

Query Parameters

Response

200 - OK. The grammar was successfully deleted from the custom language model.

400 - Bad Request. The specified customization ID or grammar name is invalid, including the case where the grammar does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for grammar name '{grammar_name}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. No grammar name was specified with the request.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

See also: Create a custom acoustic model.">

Create a custom acoustic model

POST /v1/acoustic_customizations

Creates a new custom acoustic model for a specified base model. The custom acoustic model can be used only with the base model for which it is created. The model is owned by the instance of the service whose credentials are used to create it.

You can create a maximum of 1024 custom acoustic models per owning credentials. The service returns an error if you attempt to create more than 1024 models. You do not lose any models, but you cannot create any more until your model count is below the limit.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also: Create a custom acoustic model.

Request

A recommended description of the new custom acoustic model. Use a localized description that matches the language of the custom model. Include a maximum of 128 characters in the description.

Response

The current status of the custom acoustic model:

`pending`: The model was created but is waiting either for valid training data to be added or for the service to finish analyzing added data.
`ready`: The model contains valid data and is ready to be trained. If the model contains a mix of valid and invalid resources, you need to set the `strict` parameter to `false` for the training to proceed.
`training`: The model is currently being trained.
`available`: The model is trained and ready to use.
`upgrading`: The model is currently being upgraded.
`failed`: Training of the model failed.

progress

integer

A percentage that indicates the progress of the custom acoustic model's current training. A value of `100` means that the model is fully trained. Note: The `progress` field does not currently reflect the progress of the training. The field changes from `0` to `100` when training is complete.

warnings

string

400 - Bad Request. A required parameter is null or invalid. Specific failure messages include:

`Required parameter '{name}' is missing`
`Required parameter '{name}' cannot be empty string`
`Required parameter '{name}' cannot be null`
`The base model '{model_name}' is not recognized`
`Acoustic customization is not supported for base model '{model_name}'`
`You exceeded the maximum '{model_number}' of allowed custom acoustic models. You have '{model_number}' custom acoustic models. Please remove the models you do not need or contact the IBM speech support team to apply for an exception.`

List custom acoustic models

GET /v1/acoustic_customizations

Lists information about all custom acoustic models that are owned by an instance of the service. Use the `language` parameter to see all custom acoustic models for the specified language. Omit the parameter to see all custom acoustic models for all languages. You must use credentials for the instance of the service that owns a model to list information about it.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also: Listing custom acoustic models.

Request

Path Parameters

Query Parameters

language

string

To determine the languages for which customization is available, see Language support for customization.

Response

200 - OK. The request succeeded.

customizations

array

An array of `AcousticModel` objects that provides information about each available custom acoustic model. The array is empty if the requesting credentials own no custom acoustic models (if no language is specified) or own no custom acoustic models for the specified language.

400 - Bad Request. A required parameter is null or invalid. Specific failure messages include:

`Language '{language}' is not supported for customization`

Get a custom acoustic model

GET /v1/acoustic_customizations/{customization_id}

Gets information about a specified custom acoustic model. You must use credentials for the instance of the service that owns a model to list information about it.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also: Listing custom acoustic models.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The request succeeded.

customization_id

string

string

The name of the custom acoustic model.

description

string

The description of the custom acoustic model.

base_model_name

string

The name of the language model for which the custom acoustic model was created.

status

string

The current status of the custom acoustic model:

`pending`: The model was created but is waiting either for valid training data to be added or for the service to finish analyzing added data.
`ready`: The model contains valid data and is ready to be trained. If the model contains a mix of valid and invalid resources, you need to set the `strict` parameter to `false` for the training to proceed.
`training`: The model is currently being trained.
`available`: The model is trained and ready to use.
`upgrading`: The model is currently being upgraded.
`failed`: Training of the model failed.

progress

integer

warnings

string

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

Delete a custom acoustic model

DELETE /v1/acoustic_customizations/{customization_id}

Deletes an existing custom acoustic model. The custom model cannot be deleted if another request, such as adding an audio resource to the model, is currently being processed. You must use credentials for the instance of the service that owns a model to delete it.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also: Deleting a custom acoustic model.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The custom acoustic model was successfully deleted.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials, including the case where the custom model does not exist:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

You can use the optional `custom_language_model_id` parameter to specify the GUID of a separately created custom language model that is to be used during training. Train with a custom language model if you have verbatim transcriptions of the audio files that you have added to the custom model or you have either corpora (text files) or a list of words that are relevant to the contents of the audio files. For training to succeed, both of the custom models must be based on the same version of the same base model, and the custom language model must be fully trained and available.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also:

Training failures

Training can fail to start for the following reasons:

The service is currently handling another request for the custom model, such as another training request or a request to add audio resources to the model.
The custom model contains less than 10 minutes of audio that includes speech, not silence.
The custom model contains more than 50 hours of audio (for IBM Cloud) or more that 200 hours of audio (for IBM Cloud Pak for Data). Note: For IBM Cloud, the maximum hours of audio for a custom acoustic model was reduced from 200 to 50 hours in August and September 2022. For more information, see Maximum hours of audio.
You passed a custom language model with the `custom_language_model_id` query parameter that is not in the available state. A custom language model must be fully trained and available to be used to train a custom acoustic model.
You passed an incompatible custom language model with the `custom_language_model_id` query parameter. Both custom models must be based on the same version of the same base model.
The custom model contains one or more invalid audio resources. You can correct the invalid audio resources or set the `strict` parameter to `false` to exclude the invalid resources from the training. The model must contain at least one valid resource for training to succeed.">

Train a custom acoustic model

POST /v1/acoustic_customizations/{customization_id}/train

Initiates the training of a custom acoustic model with new or changed audio resources. After adding or deleting audio resources for a custom acoustic model, use this method to begin the actual training of the model on the latest audio data. The custom acoustic model does not reflect its changed data until you train it. You must use credentials for the instance of the service that owns a model to train it.

The training method is asynchronous. Training time depends on the cumulative amount of audio data that the custom acoustic model contains and the current load on the service. When you train or retrain a model, the service uses all of the model's audio data in the training. Training a custom acoustic model takes approximately as long as the length of its cumulative audio data. For example, it takes approximately 2 hours to train a model that contains a total of 2 hours of audio. The method returns an HTTP 200 response code to indicate that the training process has begun.

You can monitor the status of the training by using the Get a custom acoustic model method to poll the model's status. Use a loop to check the status once a minute. The method returns an `AcousticModel` object that includes `status` and `progress` fields. A status of `available` indicates that the custom model is trained and ready to use. The service cannot train a model while it is handling another request for the model. The service cannot accept subsequent training requests, or requests to add new audio resources, until the existing training request completes.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also:

Training failures

Training can fail to start for the following reasons:

The service is currently handling another request for the custom model, such as another training request or a request to add audio resources to the model.
The custom model contains less than 10 minutes of audio that includes speech, not silence.
The custom model contains more than 50 hours of audio (for IBM Cloud) or more that 200 hours of audio (for IBM Cloud Pak for Data). Note: For IBM Cloud, the maximum hours of audio for a custom acoustic model was reduced from 200 to 50 hours in August and September 2022. For more information, see Maximum hours of audio.
You passed a custom language model with the `custom_language_model_id` query parameter that is not in the available state. A custom language model must be fully trained and available to be used to train a custom acoustic model.
You passed an incompatible custom language model with the `custom_language_model_id` query parameter. Both custom models must be based on the same version of the same base model.
The custom model contains one or more invalid audio resources. You can correct the invalid audio resources or set the `strict` parameter to `false` to exclude the invalid resources from the training. The model must contain at least one valid resource for training to succeed.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

custom_language_model_id

string

The customization ID (GUID) of a custom language model that is to be used during training of the custom acoustic model. Specify a custom language model that has been trained with verbatim transcriptions of the audio resources or that contains words that are relevant to the contents of the audio resources. The custom language model must be based on the same version of the same base model as the custom acoustic model, and the custom language model must be fully trained and available. The credentials specified with the request must own both custom models.

strict

boolean

If `false`, allows training of the custom acoustic model to proceed as long as the model contains at least one valid audio resource. The method returns an array of `TrainingWarning` objects that lists any invalid resources. By default (`true`), training of a custom acoustic model fails (status code 400) if the model contains one or more invalid audio resources.

Response

200 - OK. Training of the custom acoustic model started successfully.

warnings

array

400 - Bad Request. A required parameter is null or invalid, or the custom model is not ready to be trained. Specific failure messages include:

`No input data modified since last training`
`The following audio resources are invalid: '{resources}'. Fix errors before training.`
`Malformed GUID: '{customization_id}'`
`The specified custom language model '{customization_id}' is not ready for AM training and/or upgrade. Please make sure it is trained and available.`
`Failed to train. No base model version found in the catalog to match amVersion='{base_model_version}' of the acoustic custom model '{customization_id}' and lmVersion='{base_model_version}' of passed language custom model '{customization_id}'. Upgrading the acoustic custom model may help.`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Reset a custom acoustic model

POST /v1/acoustic_customizations/{customization_id}/reset

Resets a custom acoustic model by removing all audio resources from the model. Resetting a custom acoustic model initializes the model to its state when it was first created. Metadata such as the name and language of the model are preserved, but the model's audio resources are removed and must be re-created. The service cannot reset a model while it is handling another request for the model. The service cannot accept subsequent requests for the model until the existing reset request completes. You must use credentials for the instance of the service that owns a model to reset it.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also: Resetting a custom acoustic model.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The custom acoustic model was successfully reset.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

While it is being upgraded, the custom model has the status `upgrading`. When the upgrade is complete, the model resumes the status that it had prior to upgrade. The service cannot upgrade a model while it is handling another request for the model. The service cannot accept subsequent requests for the model until the existing upgrade request completes.

If the custom acoustic model was trained with a separately created custom language model, you must use the `custom_language_model_id` parameter to specify the GUID of that custom language model. The custom language model must be upgraded before the custom acoustic model can be upgraded. Omit the parameter if the custom acoustic model was not trained with a custom language model.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also: Upgrading a custom acoustic model.">

Upgrade a custom acoustic model

POST /v1/acoustic_customizations/{customization_id}/upgrade_model

Initiates the upgrade of a custom acoustic model to the latest version of its base language model. The upgrade method is asynchronous. It can take on the order of minutes or hours to complete depending on the amount of data in the custom model and the current load on the service; typically, upgrade takes approximately twice the length of the total audio contained in the custom model. A custom model must be in the `ready` or `available` state to be upgraded. You must use credentials for the instance of the service that owns a model to upgrade it.

The method returns an HTTP 200 response code to indicate that the upgrade process has begun successfully. You can monitor the status of the upgrade by using the Get a custom acoustic model method to poll the model's status. The method returns an `AcousticModel` object that includes `status` and `progress` fields. Use a loop to check the status once a minute.

While it is being upgraded, the custom model has the status `upgrading`. When the upgrade is complete, the model resumes the status that it had prior to upgrade. The service cannot upgrade a model while it is handling another request for the model. The service cannot accept subsequent requests for the model until the existing upgrade request completes.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

See also: Upgrading a custom acoustic model.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

custom_language_model_id

string

If the custom acoustic model was trained with a custom language model, the customization ID (GUID) of that custom language model. The custom language model must be upgraded before the custom acoustic model can be upgraded. The custom language model must be fully trained and available. The credentials specified with the request must own both custom models.

force

boolean

If `true`, forces the upgrade of a custom acoustic model for which no input data has been modified since it was last trained. Use this parameter only to force the upgrade of a custom acoustic model that is trained with a custom language model, and only if you receive a 400 response code and the message `No input data modified since last training`. See Upgrading a custom acoustic model.

Response

200 - OK. Upgrade of the custom acoustic model started successfully.

400 - Bad Request. A parameter is null or invalid, or the specified custom model cannot be upgraded. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Custom model is up-to-date`
`No input data available to upgrade the model`
`No input data modified since last training`
`Cannot upgrade failed custom model`
`The passed language custom model needs to be upgraded in order to upgrade the acoustic custom model.`
`The specified custom language model '{customization_id}' is not ready for AM training and/or upgrade. Please make sure it is trained and available.`
`Base model name mismatch detected. Please make sure that the base model name of the language custom model matches the base model name of the acoustic custom model.`
`Invalid model type for customization_id '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

List audio resources

GET /v1/acoustic_customizations/{customization_id}/audio

Lists information about all audio resources from a custom acoustic model. The information includes the name of the resource and information about its audio data, such as its duration. It also includes the status of the audio resource, which is important for checking the service's analysis of the resource in response to a request to add it to the custom acoustic model. You must use credentials for the instance of the service that owns a model to list its audio resources.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

Query Parameters

Response

200 - OK. The request succeeded.

total_minutes_of_audio

number

The total minutes of accumulated audio summed over all of the valid audio resources for the custom acoustic model. You can use this value to determine whether the custom model has too little or too much audio to begin training.

audio

array

An array of `AudioResource` objects that provides information about the audio resources of the custom acoustic model. The array is empty if the custom model has no audio resources.

400 - Bad Request. The specified customization ID is invalid:

`Malformed GUID: '{customization_id}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

The method is asynchronous. It can take several seconds or minutes to complete depending on the duration of the audio and, in the case of an archive file, the total number of audio files being processed. The service returns a 201 response code if the audio is valid. It then asynchronously analyzes the contents of the audio file or files and automatically extracts information about the audio such as its length, sampling rate, and encoding. You cannot submit requests to train or upgrade the model until the service's analysis of all audio resources for current requests completes.

To determine the status of the service's analysis of the audio, use the Get an audio resource method to poll the status of the audio. The method accepts the customization ID of the custom model and the name of the audio resource, and it returns the status of the resource. Use a loop to check the status of the audio every few seconds until it becomes `ok`.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

Content types for audio-type resources

You can add an individual audio file in any format that the service supports for speech recognition. For an audio-type resource, use the `Content-Type` parameter to specify the audio format (MIME type) of the audio file, including specifying the sampling rate, channels, and endianness where indicated.

`audio/alaw` (Specify the sampling rate (`rate`) of the audio.)
`audio/basic` (Use only with narrowband models.)
`audio/flac`
`audio/g729` (Use only with narrowband models.)
`audio/l16` (Specify the sampling rate (`rate`) and optionally the number of channels (`channels`) and endianness (`endianness`) of the audio.)
`audio/mp3`
`audio/mpeg`
`audio/mulaw` (Specify the sampling rate (`rate`) of the audio.)
`audio/ogg` (The service automatically detects the codec of the input audio.)
`audio/ogg;codecs=opus`
`audio/ogg;codecs=vorbis`
`audio/wav` (Provide audio with a maximum of nine channels.)
`audio/webm` (The service automatically detects the codec of the input audio.)
`audio/webm;codecs=opus`
`audio/webm;codecs=vorbis`

The sampling rate of an audio file must match the sampling rate of the base model for the custom model: for broadband models, at least 16 kHz; for narrowband models, at least 8 kHz. If the sampling rate of the audio is higher than the minimum required rate, the service down-samples the audio to the appropriate rate. If the sampling rate of the audio is lower than the minimum required rate, the service labels the audio file as `invalid`.

See also: Supported audio formats.

Content types for archive-type resources

You can add an archive file (.zip or .tar.gz file) that contains audio files in any format that the service supports for speech recognition. For an archive-type resource, use the `Content-Type` parameter to specify the media type of the archive file:

`application/zip` for a .zip file
`application/gzip` for a .tar.gz file.

When you add an archive-type resource, the `Contained-Content-Type` header is optional depending on the format of the files that you are adding:

For audio files of type `audio/alaw`, `audio/basic`, `audio/l16`, or `audio/mulaw`, you must use the `Contained-Content-Type` header to specify the format of the contained audio files. Include the `rate`, `channels`, and `endianness` parameters where necessary. In this case, all audio files contained in the archive file must have the same audio format.
For audio files of all other types, you can omit the `Contained-Content-Type` header. In this case, the audio files contained in the archive file can have any of the formats not listed in the previous bullet. The audio files do not need to have the same format.

Do not use the `Contained-Content-Type` header when adding an audio-type resource.

Naming restrictions for embedded audio files

The name of an audio file that is contained in an archive-type resource can include a maximum of 128 characters. This includes the file extension and all elements of the name (for example, slashes).">

Add an audio resource

POST /v1/acoustic_customizations/{customization_id}/audio/{audio_name}

Adds an audio resource to a custom acoustic model. Add audio content that reflects the acoustic characteristics of the audio that you plan to transcribe. You must use credentials for the instance of the service that owns a model to add an audio resource to it. Adding audio data does not affect the custom acoustic model until you train the model for the new data by using the Train a custom acoustic model method.

You can add individual audio files or an archive file that contains multiple audio files. Adding multiple audio files via a single archive file is significantly more efficient than adding each file individually. You can add audio resources in any format that the service supports for speech recognition.

You can use this method to add any number of audio resources to a custom model by calling the method once for each audio or archive file. You can add multiple different audio resources at the same time. You must add a minimum of 10 minutes of audio that includes speech, not just silence, to a custom acoustic model before you can train it. No audio resource, audio- or archive-type, can be larger than 100 MB. To add an audio resource that has the same name as an existing audio resource, set the `allow_overwrite` parameter to `true`; otherwise, the request fails. A custom model can contain no more than 50 hours of audio (for IBM Cloud) or 200 hours of audio (for IBM Cloud Pak for Data). Note: For IBM Cloud, the maximum hours of audio for a custom acoustic model was reduced from 200 to 50 hours in August and September 2022. For more information, see Maximum hours of audio.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

Content types for audio-type resources

`audio/alaw` (Specify the sampling rate (`rate`) of the audio.)
`audio/basic` (Use only with narrowband models.)
`audio/flac`
`audio/g729` (Use only with narrowband models.)
`audio/l16` (Specify the sampling rate (`rate`) and optionally the number of channels (`channels`) and endianness (`endianness`) of the audio.)
`audio/mp3`
`audio/mpeg`
`audio/mulaw` (Specify the sampling rate (`rate`) of the audio.)
`audio/ogg` (The service automatically detects the codec of the input audio.)
`audio/ogg;codecs=opus`
`audio/ogg;codecs=vorbis`
`audio/wav` (Provide audio with a maximum of nine channels.)
`audio/webm` (The service automatically detects the codec of the input audio.)
`audio/webm;codecs=opus`
`audio/webm;codecs=vorbis`

See also: Supported audio formats.

Content types for archive-type resources

`application/zip` for a .zip file
`application/gzip` for a .tar.gz file.

When you add an archive-type resource, the `Contained-Content-Type` header is optional depending on the format of the files that you are adding:

For audio files of type `audio/alaw`, `audio/basic`, `audio/l16`, or `audio/mulaw`, you must use the `Contained-Content-Type` header to specify the format of the contained audio files. Include the `rate`, `channels`, and `endianness` parameters where necessary. In this case, all audio files contained in the archive file must have the same audio format.
For audio files of all other types, you can omit the `Contained-Content-Type` header. In this case, the audio files contained in the archive file can have any of the formats not listed in the previous bullet. The audio files do not need to have the same format.

Do not use the `Contained-Content-Type` header when adding an audio-type resource.

Naming restrictions for embedded audio files

The name of an audio file that is contained in an archive-type resource can include a maximum of 128 characters. This includes the file extension and all elements of the name (for example, slashes).

Request

Request Body

The audio resource that is to be added to the custom acoustic model, an individual audio file or an archive file.

With the `curl` command, use the `--data-binary` option to upload the file for the request.

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

audio_name

Required *

string

The name of the new audio resource for the custom acoustic model. Use a localized name that matches the language of the custom model and reflects the contents of the resource.

Include a maximum of 128 characters in the name.
Do not use characters that need to be URL-encoded. For example, do not use spaces, slashes, backslashes, colons, ampersands, double quotes, plus signs, equals signs, questions marks, and so on in the name. (The service does not prevent the use of these characters. But because they must be URL-encoded wherever used, their use is strongly discouraged.)
Do not use the name of an audio resource that has already been added to the custom model.

Query Parameters

allow_overwrite

boolean

If `true`, the specified audio resource overwrites an existing audio resource with the same name. If `false`, the request fails if an audio resource with the same name already exists. The parameter has no effect if an audio resource with the same name does not already exist.

Response

201 - Created. Addition of the audio resource was successfully started. The service is analyzing the data.

400 - Bad Request. A required parameter is null or invalid, the specified customization ID or audio resource is invalid, or the specified audio resource already exists. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Audio file not specified or empty`
`Invalid audio format detected`
`Invalid or missing audio content type`
`Audio '{name}' already exists - change its name, remove existing file before adding new one, or overwrite existing file by setting 'allow_overwrite' flag to 'true'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. The audio resource name includes characters that need to be URL-encoded.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Get an audio resource

GET /v1/acoustic_customizations/{customization_id}/audio/{audio_name}

Gets information about an audio resource from a custom acoustic model. The method returns an `AudioListing` object whose fields depend on the type of audio resource that you specify with the method's `audio_name` parameter:

For an audio-type resource, the object's fields match those of an `AudioResource` object: `duration`, `name`, `details`, and `status`.
For an archive-type resource, the object includes a `container` field whose fields match those of an `AudioResource` object. It also includes an `audio` field, which contains an array of `AudioResource` objects that provides information about the audio files that are contained in the archive.

The information includes the status of the specified audio resource. The status is important for checking the service's analysis of a resource that you add to the custom model.

For an audio-type resource, the `status` field is located in the `AudioListing` object.
For an archive-type resource, the `status` field is located in the `AudioResource` object that is returned in the `container` field.

You must use credentials for the instance of the service that owns a model to list its audio resources.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

audio_name

Required *

string

The name of the audio resource for the custom acoustic model.

Query Parameters

Response

200 - OK. The request succeeded.

duration

integer

For an audio-type resource, the total seconds of audio in the resource. Omitted for an archive-type resource.

name

string

For an audio-type resource, the user-specified name of the resource. Omitted for an archive-type resource.

details

object

Information about an audio resource from a custom acoustic model.

status

string

For an audio-type resource, the status of the resource:

`ok`: The service successfully analyzed the audio data. The data can be used to train the custom model.
`being_processed`: The service is still analyzing the audio data. The service cannot accept requests to add new audio resources or to train the custom model until its analysis is complete.
`invalid`: The audio data is not valid for training the custom model (possibly because it has the wrong format or sampling rate, or because it is corrupted).

Omitted for an archive-type resource.

container

object

Information about an audio resource from a custom acoustic model.

audio

array

For an archive-type resource, an array of `AudioResource` objects that provides information about the audio-type resources that are contained in the resource. Omitted for an audio-type resource.

400 - Bad Request. The specified customization ID or audio resource name is invalid, including the case where the audio resource does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for audio name '{audio_name}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

Delete an audio resource

DELETE /v1/acoustic_customizations/{customization_id}/audio/{audio_name}

Deletes an existing audio resource from a custom acoustic model. Deleting an archive-type audio resource removes the entire archive of files. The service does not allow deletion of individual files from an archive resource.

Removing an audio resource does not affect the custom model until you train the model on its updated data by using the Train a custom acoustic model method. You can delete an existing audio resource from a model while a different resource is being added to the model. You must use credentials for the instance of the service that owns a model to delete its audio resources.

Note: Acoustic model customization is supported only for use with previous-generation models. It is not supported for large speech models and next-generation models.

Request

Path Parameters

customization_id

Required *

string

The customization ID (GUID) of the custom acoustic model that is to be used for the request. You must make the request with credentials for the instance of the service that owns the custom model.

audio_name

Required *

string

The name of the audio resource for the custom acoustic model.

Query Parameters

Response

200 - OK. The audio resource was successfully deleted from the custom acoustic model.

400 - Bad Request. The specified customization ID or audio resource name is invalid, including the case where the audio resource does not exist for the custom model. Specific failure messages include:

`Malformed GUID: '{customization_id}'`
`Invalid value for audio name '{audio_name}'`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

401 - Unauthorized. The specified credentials are invalid or the specified customization ID is invalid for the requesting credentials:

`Invalid customization_id '{customization_id}' for user`

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

405 - Method Not Allowed. No audio resource name was specified with the request.

error

string

Description of the problem.

code

integer

HTTP response code.

code_description

string

Response message.

warnings

array

Warnings associated with the error.

409 - Conflict. The service is currently busy handling a previous request for the custom model:

`Customization '{customization_id}' is currently locked to process your last request.`

Delete labeled data

DELETE /v1/user_data

Deletes all data that is associated with a specified customer ID. The method deletes all data for the customer ID, regardless of the method by which the information was added. The method has no effect if no data is associated with the customer ID. You must issue the request with credentials for the same instance of the service that was used to associate the customer ID with the data. You associate a customer ID with data by passing the `X-Watson-Metadata` header with a request that passes the data.

Note: If you delete an instance of the service from the service console, all data associated with that service instance is automatically deleted. This includes all custom language models, corpora, grammars, and words; all custom acoustic models and audio resources; all registered endpoints for the asynchronous HTTP interface; and all data related to speech recognition requests.

See also: Information security.

Request

Path Parameters

Query Parameters

customer_id

Required *

string

The customer ID for which all data is to be deleted.

Response

200 - OK. The deletion request was successfully submitted.

400 - Bad Request. The request did not pass a customer ID:

`No customer ID found in the request`

POST /v1/detect_language

Detects the spoken language in audio streams. The endpoint is `/v1/detect_language` and user can optionally include `lid_confidence` parameter to set a custom confidence threshold for detection. The model continuously processes incoming audio and returns the identified language when it reaches a confidence level higher than the specified threshold (0.99 by default). See Spoken language identification.

Request

Request Body

Path Parameters

Query Parameters

lid_confidence

Required *

number

Set a custom confidence threshold for detection.

Response

200 - OK. The request succeeded.

results

array

An array of `LanguageDetectionResult` objects

result_index

integer