Release notes for Text to Speech for IBM Cloud

The service now supports a new female expressive neural voice for Latin American Spanish

• es-LA_DanielaExpressive.

Expressive neural voices offer natural-sounding speech that is clear, crisp, and fluid. The new voice is generally available (GA) for production use. It supports the use of both the standard International Phonetic Alphabet (IPA) and IBM Symbolic Phonetic Representation (SPR) phonetic symbols. For more information, see

• Expressive neural voices.
• Spanish expressive voices.
• Spanish symbols.

Important: All neural voices have reached their end of service date and have been removed from the service and the documentation. No enhanced neural or expressive neural voices are affected. For a complete list of all obsolete neural voices, see the 1 March 2023 service updates. An attempt to use an obsolete voice returns the HTTP response code 404 with the message Model '{voice}' not found.

New enhanced neural and expressive neural voices are available for the Australian English, Korean, and Netherlands Dutch languages. You must migrate to one of the new voices for Australian English, Korean, or Netherlands Dutch to continue by using the obsolete languages.

For more information, see

• Languages and voices
• Migrating from neural voices

The service now supports a new enhanced neural female voice for Netherlands Dutch: nl-NL_MerelV3Voice. It supports the use of both the standard International Phonetic Alphabet (IPA) and IBM Symbolic Phonetic Representation (SPR) phonetic symbols.

The new voice is beta functions pending completion of support for SSML. At its initial release, the voice does not support use of the following SSML-related functions:

• The <prosody> element with any speech synthesis request
• The rate_percentage and pitch_percentage parameters with any speech synthesis request
• The <mark> element with a WebSocket speech synthesis request
• The timings parameter of the JSON text message with a WebSocket speech synthesis request

For more information about the new voice, its support for IPA and SPR symbols, and migrating to the new voice from the deprecated Netherlands Dutch neural voices, see

• Enhanced neural voices
• Dutch (Netherlands) symbols
• Migrating from neural voices

Defect fix: In the documentation for Korean SPR symbols, two-character symbols for consonants are now enclosed in single quotation marks, making them a single symbol. Previously, they were shown as two separate symbols, without enclosing quotation marks. For more information, see Consonants (Korean).

The overview documentation for IBM SPR symbols has been updated to clarify the use of multi-character symbols. For more information, see Speech sound symbols).

Important: Effective 31 March 2023, all neural voices reach their end of service date and will be removed from the service and the documentation. The neural voices have been deprecated since 31 March 2022. No enhanced neural or expressive neural voices are effected.

The following neural voices are removed:

• Arabic: ar-MS_OmarVoice
• Chinese (Mandarin): zh-CN_LiNaVoice, zh-CN_WangWeiVoice, and zh-CN_ZhangJingVoice
• Czech: cs-CZ_AlenaVoice
• Dutch (Belgian): nl-BE_AdeleVoice and nl-BE_BramVoice
• Dutch (Netherlands): nl-NL_EmmaVoice and nl-NL_LiamVoice
• English (Australian): en-AU_CraigVoice, en-AU_MadisonVoice, and en-AU_SteveVoice
• Korean: ko-KR_HyunjunVoice, ko-KR_SiWooVoice, ko-KR_YoungmiVoice, and ko-KR_YunaVoice
• Swedish: sv-SE_IngridVoice

New enhanced neural and expressive neural voices are already available for the Australian English and Korean languages. In the coming weeks, a new enhanced neural voice is available for the Netherlands Dutch language. You must migrate to one of the new voices for Australian English, Korean, or Netherlands Dutch before 31 March 2023.

For more information, see

• Languages and voices
• Migrating from neural voices

The service now supports two new expressive neural voices, male and female, for Australian English:

• en-AU_HeidiExpressive
• en-AU_JackExpressive

Expressive neural voices offer natural-sounding speech that is exceptionally clear, crisp, and fluid. The new voices are generally available (GA) for production use. They support the use of both the standard International Phonetic Alphabet (IPA) and IBM Symbolic Phonetic Representation (SPR) phonetic symbols. For more information, see

• Expressive neural voices
• English (Australian) symbols

You can migrate from Australian English neural voices that are deprecated to the new expressive neural voices. For more information, see

• Migrating from neural voices

The service now supports a new enhanced neural female voice for Korean: ko-KR_JinV3Voice. The new voice is generally available (GA) for production use. It supports the use of both the standard International Phonetic Alphabet (IPA) and IBM Symbolic Phonetic Representation (SPR) phonetic symbols. For more information, see

• Enhanced neural voices
• Korean symbols

You can migrate from Korean neural voices that are deprecated to the new enhanced neural voice. For more information, see

• Migrating from neural voices

Defect fix: The French and Canadian voices now pronounce times like 19:41 correctly. Previously, the voices were omitting elements of the time in the synthesized audio.

Defect fix: The Japanese voice no longer inserts unexpected audio in speech synthesis results. Previously, other audio was inserted in certain cases.

IBM announced the deprecation of IBM Cloud Foundry on 31 May 2022. As of 30 November 2022, new IBM Cloud Foundry applications cannot be created and only existing users are able to deploy applications. IBM Cloud Foundry reaches end of support on 1 June 2023. Then, any IBM Cloud Foundry application runtime instances running IBM Cloud Foundry applications will be permanently disabled, deprovisioned, and deleted.

To continue to use your IBM Cloud applications beyond 1 June 2023, you must migrate to resource groups before that date. Resource groups are conceptually similar to Cloud Foundry spaces. They include several extra benefits, such as finer-grained access control by using IBM Cloud Identity and Access Management (IAM), the ability to connect service instances to apps and services across different regions, and an easy way to view usage per group.

Defect fix: You can now use the <say-as> element to pronounce large numbers as cardinal numbers. Previously, enclosing a large number in the <say-as> element with the attribute interpret-as="cardinal" might cause speech synthesis to fail for English voices. For example, <say-as interpret-as="cardinal">3,200</say-as> might cause the service to generate an error. For more information, see cardinal in the topic SSML elements.

Defect fix: The service now pronounces homonyms and other words correctly based on their context in English text that is to be synthesized. Previously, words such as advocate and wifi might be pronounced incorrectly by English voices.

• Creating a custom model
• API and SDK reference

The service offers four new expressive neural voices for US English:

• en-US_AllisonExpressive
• en-US_EmmaExpressive
• en-US_LisaExpressive
• en-US_MichaelExpressive

Expressive neural voices offer natural-sounding speech that is exceptionally clear, crisp, and fluid. The new voices are generally available (GA) for production use. They support the use of both the standard International Phonetic Alphabet (IPA) and IBM Symbolic Phonetic Representation (SPR) phonetic symbols. For more information, see

• Expressive neural voices
• English (United States) symbols

The expressive neural voices determine the sentiment of the text from the context of its words and phrases. The speech that they produce, in addition to having a conversational style, reflects the mood of the text. But you can embellish the voices' natural tendencies by indicating that all or some of the text is to emphasize one of the following speaking styles:

• Cheerful - Expresses happiness and good news.
• Empathetic - Expresses empathy or sympathy.
• Neutral - Expresses objectivity and evenness.
• Uncertain - Expresses confusion or uncertainty.

For more information, see Using speaking styles.

With expressive neural voices, the service automatically detects a set of common interjections based on context. When it synthesizes these interjections, it gives them the natural emphasis that a human would use in normal conversation. For some of the interjections, you can use SSML to enable or disable their emphasis. For more information, see Emphasizing interjections.

The expressive voices use a conversational style that naturally applies the correct intonation from context. But you can indicate that one or more words are to be given more or less emphasis. The change in stress can be indicated by an increase or decrease in pitch, timing, volume, or other acoustic attributes. For more information, see Emphasizing words.

• For more information about the available formats, see Supported audio formats.
• For more information about specifying a format, see Specifying an audio format.

Important: Effective 31 March 2022, all neural voices are deprecated. The deprecated voices remain available to existing users until 31 March 2023, when they will be removed from the service and the documentation. No enhanced neural voices or expressive voices are deprecated.

The following neural voices are now deprecated:

• Arabic: ar-MS_OmarVoice
• Chinese (Mandarin): zh-CN_LiNaVoice, zh-CN_WangWeiVoice, and zh-CN_ZhangJingVoice
• Czech: cs-CZ_AlenaVoice
• Dutch (Belgian): nl-BE_AdeleVoice and nl-BE_BramVoice
• Dutch (Netherlands): nl-NL_EmmaVoice and nl-NL_LiamVoice
• English (Australian): en-AU_CraigVoice, en-AU_MadisonVoice, and en-AU_SteveVoice
• Korean: ko-KR_HyunjunVoice, ko-KR_SiWooVoice, ko-KR_YoungmiVoice, and ko-KR_YunaVoice
• Swedish: sv-SE_IngridVoice

The deprecated neural voices continue to be available to existing users but are no longer available to new users:

• Existing users: Instances of the service created before 31 March 2022 can continue to use the deprecated voices for speech synthesis. The instances can also be used to create and work with custom models that are based on the deprecated voices. And they can continue to list and query the voices with the GET /v1/voices and GET /v1/voices/{voice} methods.
• New users: Instances of the service that is created on or after 31 March 2022 cannot use the deprecated voices for speech synthesis. The instances also cannot be used to create custom models that are based on the deprecated voices. And they cannot list or query the voices with the GET /v1/voices and GET /v1/voices/{voice} methods. Any API call that includes one of the deprecated voices returns an HTTP error code of 400 or 404, depending on the call.

New voices for Australian English, Dutch, and Korean will be released by 15 February 2023. If you are using Australian English, Dutch, or Korean voices, your API calls will automatically redirect to the new voices in that language. Voices in Arabic, Chinese, Czech, Swedish, and Flemish will be removed from service.

For more information about all available languages and voices, see Languages and voices.

The standard concatenating voices were deprecated on 2 December 2020. These standard voices have now been removed from the API reference. The topic Migrating from standard to neural voices has also been removed from the page Languages and voices.

Similarly, the former name of the Arabic voice, ar-AR_OmarVoice, was deprecated at the same time. It too has been removed from the documentation. Use the voice ar-MS_OmarVoice instead.

The response object that the service sends when you request word timings with the WebSocket interface have changed. The service now sends word timing results in a single array that includes a string followed by two floats:

{
  "words": [
    ["Hello", 0.0, 0.259],
    ["world", 0.259, 0.532]
  ]
}

The service previously sent timing results as an array that included a string following by an array of two floats:

{
  "words": [
    ["Hello", [0.0629826778195474, 0.2590192737303819]],
    ["world", [0.2598829173456253, 0.5322130804452672]]
  ]
}

Also, the level of precision for word timings and marks is now reduced to three decimal places. For more information about the new responses, see Generating word timings.

Note: Results for enhanced neural and neural voices were different previously. These inconsistencies might cause errors for the Watson SDKs. The results for all voices are now consistent.

• The examples of the <break> element are now correct. The element is unary, as now shown in the examples. The previous examples included open and close tags with embedded text. The embedded text was not spoken by the service. For more information, see The <break> element.
• The service supports SSML version 1.1. All references and examples now use the correct version. The documentation previously referred to version 1.0.

A new language, Czech, with a new female voice, cs-CZ_AlenaVoice, is now available. The voice is a neural voice.

A new male Belgian Dutch (Flemish) voice, nl-BE_BramVoice, is now available. The voice is a neural voice.

Defect fix: The following defects for the Speech Synthesis Markup Language (SSML) and speech synthesis were fixed with this release:

• The pitch attribute of the <prosody> element is now applied to all specified text. Previously, the pitch change was not always applied to the first word of the affected text. Also, the documentation now includes more guidance about specifying a pitch value. For more information, see The pitch attribute.
• Speech synthesis of Japanese text now speaks the audio more slowly. Previously, the synthesized speech was being spoken too quickly. If you find that synthesis of Japanese text is still spoken too quickly for your application, use the rate attribute of the SSML <prosody> element to control the rate of speech. For more information, see The rate attribute.
• Neural voices now parse the escaped apostrophe character (') properly. Previously, some neural voices were not interpreting the character properly.

The service now supports Canadian French with the enhanced neural voice fr-CA_LouiseV3Voice. The Canadian French voice supports customization and is generally available (GA) for production use.

• To hear a sample of the new voice, see Supported languages and voices.
• For more information about the phonetic symbols and Unicode values that are available for the Canadian French language, see French (Canadian) symbols.

The new Tune by Example feature lets you control how specified text is spoken by the service. The feature is beta functions that is supported only for US English custom models and voices. It has two components:

• Custom prompts include the written text that is to be spoken and recorded audio that speaks the text as you want to hear it. The audio specifies the intonation, cadence, and stress of the synthesized text. The prompt can emphasize different syllables or words, introduce pauses, and generally make the synthesized audio sound more natural and appropriate for its context.
• Speaker models provide enrollment audio for a user who speaks one or more prompts. A speaker model provides an audio sample of a user's voice. The service trains itself on the voice, which can help it to produce higher-quality prompts for that speaker.

You specify a custom prompt with a speech synthesis request to indicate how the service's voice is to pronounce the text. To specify a prompt, you use the SSML extension <ibm:prompt id="{prompt_id}"/>. The synthesized audio duplicates the prosody of the prompt.

For more information about using the Tune by Example feature, see the following topics:

• Understanding Tune by Example
• Rules for creating custom prompts and speaker models
• Creating a custom prompt
• Using a custom prompt for speech synthesis
• Managing custom prompts
• Managing speaker models

The service includes eight new methods for working with the Tune by Example feature. The descriptions of the new methods that follow provide links to their entries in the API and SDK reference. You might need to select the Curl tab of the reference to see the new methods.

• The service includes four methods for working with custom prompts:

  • Add a custom prompt: POST /v1/customizations/{customization_id}/prompts/{prompt_id}
  • List custom prompts: GET /v1/customizations/{customization_id}/prompts
  • Get a custom prompt: GET /v1/customizations/{customization_id}/prompts/{prompt_id}
  • Delete a custom prompt: DELETE /v1/customizations/{customization_id}/prompts/{prompt_id}

• The service includes four methods for working with speaker models:

  • Create a speaker model: POST /v1/speakers
  • List speaker models: GET /v1/speakers
  • Get a speaker model: GET /v1/speakers/{speaker_id}
  • Delete a speaker model: DELETE /v1/speakers/{speaker_id}

Activity Tracker actions are available for all new Tune by Example events. For more information, see Tune by Example events.

The names of the actions for the Activity Tracker events for the customization methods have changed. The actions now include the string custom-model instead of custom-voice. The old names of the actions are deprecated. The old names are still available for use but will be removed at a future date. Migrate to the new names that are listed in Customization events at your earliest convenience.

Create events Deprecated action name -> New action name

• text-to-speech.custom-voice.create -> text-to-speech.custom-model.create
• text-to-speech.custom-voice-word-list.create -> text-to-speech.custom-model-word-list.create
• text-to-speech.custom-voice-word.create -> text-to-speech.custom-model-word.create

Read events Deprecated action name -> New action name

• text-to-speech.custom-voice-list.read -> text-to-speech.custom-model-list.read
• text-to-speech.custom-voice.read -> text-to-speech.custom-model.read
• text-to-speech.custom-voice-word-list.read -> text-to-speech.custom-model-word-list.read
• text-to-speech.custom-voice-word.read -> text-to-speech.custom-model-word.read

Update event Deprecated action name -> New action name

• text-to-speech.custom-voice.update -> text-to-speech.custom-model.update

Delete events Deprecated action name -> New action name

• text-to-speech.custom-voice.delete -> text-to-speech.custom-model.delete
• text-to-speech.custom-voice-word.delete -> text-to-speech.custom-model-word.delete

The voices that are offered by the service have undergone significant change. The service supports new languages and voices, has improved the quality of many voices, and has deprecated many older voices. In addition, all of the services voices are now customizable and generally available (GA) for production use.

To optimize the overall quality of voice synthesis, all available voices are now based on neural technology. The service offers two types of voices that are based on neural technology:

• Neural voices, which do not include the string V3 in their names, are now available for Arabic, Australian English, Chinese, Netherlands Dutch, and Korean. Neural voices support the use of the International Phonetic Alphabet (IPA) with the Speech Synthesis Markup Language (SSML) <phoneme> element.
• Enhanced neural voices, which include the string V3 in their names, are now available for Brazilian Portuguese, United Kingdom and United States English, French, German, Italian, Japanese, and Spanish (all dialects). Enhanced neural voices support the use of both IPA and IBM Symbolic Phonetic Representation (SPR) with the SSML <phoneme> element. Enhanced neural voices also achieve a slightly higher degree of natural-sounding speech. And further customization capabilities are exposed for enhanced neural voices in the coming months.

The service no longer offers standard voices for any language. Standard voices used concatenating synthesis to assemble segments of recorded speech to generate the requested audio.

For more information, see Languages and voices.

The following Australian English and Korean voices are new:

• The service now supports Australian English with the following two neural voices: en-AU_CraigVoice and en-AU_MadisonVoice.
• The service now supports two new Korean neural voices: ko-KR_HyunjunVoice and ko-KR_SiWooVoice.

The voices for the existing Arabic, Chinese, Netherlands Dutch, and Korean languages, all of which were concatenating, are now neural:

• ar-MS_OmarVoice
• ko-KR_YoungmiVoice
• ko-KR_YunaVoice
• nl-NL_EmmaVoice
• nl-NL_LiamVoice
• zh-CN_LiNaVoice
• zh-CN_WangWeiVoice
• zh-CN_ZhangJingVoice

The following other changes have also been made:

• The Arabic voice is now named ar-MS_OmarVoice. The former name, ar-AR_OmarVoice, is deprecated. It will continue to function for at least one year but might be removed at a future date. You are encouraged to migrate to the new name at your earliest convenience.
• The Arabic language now supports the use of IPA symbols and Unicode values with the SSML <phoneme> element. To create a custom model for Arabic, you must use the language identifier ar-MS. The identifier ar-AR is not supported for customization.
• The IPA symbols for the Arabic language are new. The previously documented symbols have been replaced.
• The IPA symbols for the Netherlands Dutch language have been changed as follows:
  • Netherlands Dutch no longer supports the following IPA symbols: tʲ (0074+02B2), ɲ (0272), ʦ (02A6), and ʔ (0294).
  • Netherlands Dutch now supports the following IPA symbol: ɣ (0263).

The following standard concatenating voices are now deprecated:

• de-DE_BirgitVoice
• de-DE_DieterVoice
• en-GB_KateVoice
• en-US_AllisonVoice
• en-US_LisaVoice
• en-US_MichaelVoice
• es-ES_EnriqueVoice
• es-ES_LauraVoice
• es-LA_SofiaVoice
• es-US_SofiaVoice
• fr-FR_ReneeVoice
• it-IT_FrancescaVoice
• ja-JP_EmiVoice
• pt-BR_IsabelaVoice

All of the standard voices that have been deprecated have equivalent neural counterparts, so no voice is being taken away. Rather, you can change to the equivalent neural version of the voice (for example, from de-DE_BirgitVoice to de-DE_BirgitV3Voice) for better speech-synthesis results.

These deprecated voices are removed from the published documentation. They will continue to function for at least one year but might be removed at a future date. You are encouraged to migrate to the equivalent neural voices at your earliest convenience.

If you omit the optional voice parameter from a speech synthesis request, the service uses en-US_MichaelV3Voice by default. This neural voice replaces the now-deprecated en-US_MichaelVoice standard voice that was the previous default.

The following features were available only for standard concatenating voices. They are deprecated and have been removed from the published documentation. They will continue to function for at least one year but might be removed at a future date. You are encouraged to remove these features from your applications and to migrate to neural voices at your earliest convenience.

• Expressive SSML. This was an IBM extension to SSML that was supported only for the en-US_AllisonVoice voice.
• Voice transformation SSML. This was an IBM extension to SSML that was supported only for the en-US_AllisonVoice, en-US_LisaVoice, and en-US_MichaelVoice voices.
• The volume attribute of the <prosody> element. This attribute was available for all standard voices.

Including these SSML elements with a synthesis request for a neural voice generates an HTTP 400 response code because the request fails SSML validation. For more information about SSML validation, see SSML validation.

Cross-Origin Resource Sharing (CORS) support is now available for all voices from the Google Chrome™ and Apple® Safari browsers. It is not available from the Mozilla Firefox™ browser for voices in the following languages: Arabic, Australian English, Chinese, Netherlands Dutch, and Korean. For more information, see using CORS support.

Defect fix: For the ja-JP_EmiV3Voice voice, the service now correctly parses SSML input text that includes a prosody rate specification. Previously, the following use of the <prosody> element worked properly:

<speak>成功する/繁栄する</speak>

But the following use of the rate attribute with the <prosody> element caused the service to read and speak the embedded SSML notation:

<speak>
  <prosody rate="fast">成功する/繁栄する</prosody>
</speak>

The service now correctly parses and applies the rate attribute of the <prosody> element for Japanese input.

The service now offers three new neural voices:

• UK English: en-GB_CharlotteV3Voice and en-GB_JamesV3Voice
• French: fr-FR_NicolasV3Voice

It also offers an improved version of the existing UK neural voice, en-GB_KateV3Voice. For more information about all available voices, see Languages and voices.

The service now supports the digits attribute of the SSML <say-as> element with its Japanese voice. For more information, see The say-as element.

The service now supports two standard female Korean voices: ko-KR_YoungmiVoice and ko-KR_YunaVoice. The following information applies to both Korean voices:

• The voices are beta functions. They might not be ready for production use and are subject to change. They are initial offerings that are expected to improve in quality with time and usage.
• The voices support customization and the /v1/pronunciation method.
• The voices support the <mark> element and the timings parameter that are available with the WebSocket interface.
• The voices support all Speech Synthesis Markup Language (SSML) elements except for expressive SSML and voice transformation SSML.
• The voices support only the International Phonetic Alphabet (IPA), not IBM Symbolic Phonetic Representation (SPR), with the <phoneme> element.

The beta Arabic, Chinese, and Netherlands Dutch voices now support the following features:

• Customization and the /v1/pronunciation method.
• The <mark> element and the timings parameter that are available with the WebSocket interface.

See the following sections for additional information:

• For more information about all available voices, see Languages and voices.
• For more information about using the WebSocket interface to obtain word timings, see Generating word timings.
• For more information about customization, see Understanding customization.
• For more information about the use of IPA and SPR with customization, see Understanding phonetic symbols. (The IPA symbols for the Arabic, Chinese, Netherlands Dutch, and Korean languages are not yet documented. This documentation will be made available soon.)

The service now supports five new neural voices:

• US English: en-US_EmilyV3Voice, en-US_HenryV3Voice, en-US_KevinV3Voice, and en-US_OliviaV3Voice
• German: de-DE_ErikaV3Voice

The new voices do not support the following SSML elements:

• Expressive SSML with the <express-as> element
• Voice Transformation with the <voice-transformation> element
• The volume attribute of the <prosody> element

For more information about these and all available voices, see Languages and voices.

The service now supports the use of Activity Tracker events for all customization operations. IBM Cloud Activity Tracker records user-initiated activities that change the state of a service in IBM Cloud. You can use this service to investigate abnormal activity and critical actions and to comply with regulatory audit requirements. In addition, you can be alerted about actions as they happen. For more information, see Activity Tracker events.

The service now supports six new standard voices in three new languages:

• Arabic: ar-AR_OmarVoice
• Chinese (Mandarin): zh-CN_LiNaVoice, zh-CN_WangWeiVoice, and zh-CN_ZhangJingVoice
• Netherlands Dutch: nl-NL_EmmaVoice and nl-NL_LiamVoice

The following information applies to these new standard voices:

• The new voices are beta functions. The voices might not be ready for production use and are subject to change. They are initial offerings that are expected to improve in quality with time and usage.
• The voices do not support the <mark> element and timings parameter that are available with the WebSocket interface.
• The voices do not support customization or the /v1/pronunciation method.
• The voices do not support expressive SSML or voice transformation SSML.
• The voices do support all other SSML elements. However, they support only the International Phonetic Alphabet (IPA), not IBM Symbolic Phonetic Representation (SPR), with the <phoneme> element.

For more information about these and all available voices, see Languages and voices.

The Text to Speech service now supports the full implementation of IBM Cloud Identity and Access Management (IAM). API keys for Watson services are no longer limited to a single service instance. You can create access policies and API keys that apply to more than one service, and you can grant access between services. For more information about IAM, see Authenticating to Watson services.

To support this change, the API service endpoints use a different domain and include the service instance ID. The pattern is api.{location}.text-to-speech.watson.cloud.ibm.com/instances/{instance_id}.

• Example HTTP URL for an instance hosted in the Dallas location:

  https://api.us-south.text-to-speech.watson.cloud.ibm.com/instances/6bbda3b3-d572-45e1-8c54-22d6ed9e52c2

• Example WebSocket URL for an instance hosted in the Dallas location:

  wss://api.us-south.text-to-speech.watson.cloud.ibm.com/instances/6bbda3b3-d572-45e1-8c54-22d6ed9e52c2

For more information about the URLs, see the API and SDK reference.

These URLs do not constitute a breaking change. The new URLs work for both your existing service instances and for new instances. The original URLs continue to work on your existing service instances for at least one year, until December 2020.

Support for the following new network and data security features is now available:

• Support for private network endpoints

  Users of Premium plans can create private network endpoints to connect to the Text to Speech service over a private network. Connections to private network endpoints do not require public internet access. For more information, see Public and private network endpoints.

• Support for data encryption with customer-managed keys

  Users of new Premium and Dedicated instances can integrate IBM® Key Protect for IBM Cloud® with the Text to Speech service to encrypt your data and manage encryption keys. For more information, see Protecting sensitive information in your Watson service.

The service now offers two versions of most of its available voices:

• Standard voices that use concatenating synthesis to assemble segments of recorded speech to generate audio. Standard voices do not include a version string in their name (for example, en-US_AllisonVoice).
• Neural voices that use Deep Neural Networks (DNNs) to predict the acoustic (spectral) features of the speech. Neural voices include a version string (V3) in their name (for example, en-US_AllisonV3Voice).

Neural versions are available for all standard voices except for the ja-JP_EmiVoice voice, which is pending and will be available soon. You cannot use the SSML <express-as> and <voice-transformation> elements with the neural voices, and you cannot use the volume attribute of the <prosody> element with the neural voices. For more information about all available voices, see Languages and voices.

The service no longer includes the V2 DNN voices that were previously available. If you use a V2 voice in your application, the service automatically uses the equivalent V3 voice instead.

The service now offers V2 Deep Neural Network (DNN) versions of its German voices:

• de-DE_BirgitV2Voice
• de-DE_DieterV2Voice

For more information about DNN-based voices, see Languages and voices.

All of the service's DNN-based voices now support the pitch and rate attributes of the SSML <prosody> element. The DNN-based voices do not support the volume attribute of the <prosody> element. For more information, see The prosody element.

Users can now see only service credential information that is associated with the role that has been assigned to their IBM Cloud account. For example, if you are assigned a reader role, any writer or higher levels of service credentials are no longer visible.

This change does not affect API access for users or applications with existing service credentials. The change affects only the viewing of credentials within IBM Cloud.

The service now offers four new V2 voices that use deep-learning synthesis to generate audio:

• en-US_AllisonV2Voice
• en-US_LisaV2Voice
• en-US_MichaelV2Voice
• it-IT_FrancescaV2Voice

These new voices use machine learning and a DNN to synthesize text to speech. Deep-learning, or Deep Neural Network (DNN)-based, synthesis produces audio with a more natural prosody and a more consistent overall quality.

But the new voices also produce audio with different signal qualities from the existing voices, so they might not be appropriate for all applications. Also, the new voices do not support the SSML elements <prosody>, <express-as>, and <voice-transformation>.

For more information about these DNN-based voices and how they differ from the existing voices, see Languages and voices.

The WebSocket interface now supports token-based Identity and Access Management (IAM) authentication from browser-based JavaScript code. The limitation to the contrary has been removed. To establish an authenticated connection with the WebSocket /v1/synthesize method:

• If you use IAM authentication, include the access_token query parameter.
• If you use Cloud Foundry service credentials, include the watson-token query parameter.

For more information, see Open a connection.

The Text to Speech service has migrated to token-based Identity and Access Management (IAM) authentication for all locations. All IBM Cloud services now use IAM authentication. The Text to Speech service migrated in each location on the following dates:

• Dallas (us-south): October 30, 2018
• Frankfurt (eu-de): October 30, 2018
• Washington, DC (us-east): June 12, 2018
• Sydney (au-syd): May 15, 2018

The migration to IAM authentication affects new and existing service instances differently:

• All new service instances that you create in any location now use IAM authentication to access the service. You can pass either a bearer token or an API key: Tokens support authenticated requests without embedding service credentials in every call; API keys use HTTP basic authentication. When you use any of the Watson SDKs, you can pass the API key and let the SDK manage the lifecycle of the tokens.
• Existing service instances that you created in a location before the indicated migration date continue to use the {username} and {password} from their previous Cloud Foundry service credentials for authentication until you migrate them to use IAM authentication.

For more information, see the following documentation:

• To learn which authentication mechanism your service instance uses, view your service credentials by clicking the instance on the IBM Cloud dashboard.
• For more information about using IAM tokens with Watson services, see Authenticating to Watson services.
• For examples that use IAM authentication, see the API & SDK reference.

The following features are enabled for applications that are hosted in Washington, DC (us-east):

• The service now supports a new API authentication process. For more information, see the 30 October 2018 service update.
• The service now supports the X-Watson-Metadata header and the DELETE /v1/user_data method. For more information, see Information security.

The following features are enabled for applications that are hosted in Sydney (au-syd):

• The service now supports a new API authentication process. For more information, see the 30 October 2018 service update.
• The service now supports the X-Watson-Metadata header and the DELETE /v1/user_data method. For more information, see Information security.

• The POST /v1/customizations method now returns 201 (instead of 200).
• The POST /v1/customizations/{customization_id} method now returns 200 (instead of 201).
• The POST /v1/customizations/{customization_id}/words method now returns 200 (instead of 201).
• The PUT /v1/customizations/{customization_id}/words/{word} method now returns 200 (instead of 201).

The service includes a new voice, es-LA_SofiaVoice, which is the Latin American equivalent of the es-US_SofiaVoice voice. The most significant difference between the two voices concerns how they interpret a $ (dollar sign): The Latin American version uses the term pesos; the North American version uses the term dolares. Other minor differences might also exist between the two voices.

In addition to the en-US_AllisonVoice, two more voices are now transformable with SSML voice transformation: en-US_LisaVoice and en-US_MichaelVoice.

When you use the customization interface with Japanese, the service now matches the longest word from the word/translation pairs that are defined for a custom model. For example, consider the following two entries for a custom model:

{
  "words": [
    {"word":"ＮＹ", "translation":"ニューヨーク", "part_of_speech":"Mesi"},
    {"word":"ＮＹＣ", "translation":"ニューヨークシティ", "part_of_speech":"Mesi"}
  ]
}

If the service finds the string ＮＹＣ in the input text, it matches that word because it is a longer match than ＮＹ. Previously, the service would have matched the string ＮＹ. For more information about working with Japanese entries for a custom model, see Working with Japanese entries.

The service now offers a WebSocket interface for synthesizing text to speech. The interface offers the same features as the /v1/synthesize method of the HTTP interface. It accepts plain text or text that is marked up with SSML. In addition, it also supports use of the SSML <mark> element to identify the time in the audio at which it finishes synthesizing all text that precedes the mark. For more information, see The WebSocket interface.

The service now offers support for text that is annotated with SSML for the languages Castilian and North American Spanish, Italian, and Brazilian Portuguese. The service already supported the use of SSML for US and UK English, French, and German. As of this update, the service supports SSML for all languages but Japanese. Moreover, you can use both IBM SPR and IPA notations to define word pronunciations with the SSML <phoneme> element. For more information, see Understanding SSML and Understanding phonetic symbols.

For US English, you can also use the SSML <phoneme> element to create word entries in a custom model; customization is supported only for US English. For more information, see Understanding customization.

The service features improved expressiveness and naturalness for the most frequently used voices. The improvements are based on Recursive Neural Network (RNN)-based prosody prediction from input text. They are made available as a new service engine and voice-model updates for the following languages:

• en-US_AllisonVoice
• en-US_LisaVoice
• en-US_MichaelVoice
• es-ES_EnriqueVoice
• fr-FR_ReneeVoice

The GET /v1/pronunciation method now accepts an optional customization_id query parameter. The parameter obtains a word translation from a specified custom model. If the custom model does not contain the word, the method returns the word's default pronunciation. For more information, see the API and SDK reference.

When using the GET /v1/pronunciation method without a customization ID and for a language other than US English, you can request a word's pronunciation only in IBM SPR notation. For a language other than US English, you must specify spr with the method's format option.

The list of supported audio formats now includes audio/basic, which provides single-channel audio that is encoded using 8-bit u-law (or mu-law) data that is sampled at 8 kHz. For more information, see Using audio formats.

The HTTP and WebSocket /v1/synthesize methods can return a warnings response that includes messages about invalid query parameters or JSON fields that are included with a request. The format of the warnings changed. The following example shows the previous format:

"warnings": "Unknown arguments: [u'{invalid_arg_1}', u'{invalid_arg_2}']."

The same warning now has the following format:

"warnings": "Unknown arguments: {invalid_arg_1}, {invalid_arg_2}."

The service offers a new customization interface that you can use to specify how it pronounces unusual words that occur in your input. The interface includes some new methods that you can use to create and manage custom models and the word/translation pairs that they contain. You can then use your custom models when synthesizing text to audio.

The service supports sounds-like translations and phonetic translations. Phonetic translations can use either the standard International Phonetic Alphabet (IPA) representation or the proprietary IBM Symbolic Phonetic Representation (SPR). You use SSML to specify phonetic translations.

The customization interface includes a collection of new HTTP methods that have the names POST /v1/customizations, POST /v1/customizations/{customization_id}, POST /v1/customizations/{customization_id}/words, and PUT /v1/customizations/{customization_id}/words/{word}. The service also provides a new GET /v1/pronunciation method that returns the pronunciation for any word and a new GET /v1/voices/{voice} method that returns detailed information about a specific voice. In addition, existing methods of the service's interface now accept custom model parameters as needed.

For more information about customization and its interface, see Understanding customization and the API and SDK reference.

The customization interface is a beta release that currently supports US English only. All customization methods and the GET /v1/pronunciation method can currently be used to create and manipulate custom models and word translations only in US English.

The service supports a new voice, pt-BR_IsabelaVoice, to synthesize audio in Brazilian Portuguese with a female voice. For more information, see Languages and voices.

Two new beta mobile Software Development Kits (SDKs) are available for the speech services. The SDKs enable mobile applications to interact with both the Text to Speech and Speech to Text services. You can use the SDKs to send text to the Text to Speech service and receive an audio response.

• The Watson Swift SDK is available from the swift-sdk repository in the watson-developer-cloud namespace on GitHub.
• The Watson Speech Android™ SDK is available from the speech-android-sdk repository in the watson-developer-cloud namespace on GitHub.

Both SDKs support authenticating with the speech services by using either your IBM Cloud service credentials or an authentication token. Because the SDKs are beta functions, they are subject to change in the future.

The service supports a new language, Japanese. The voice ja-JP_EmiVoice is a Japanese female voice.

The service moved from beta to general availability (GA) on 1 July 2015. The following differences that are existed between the beta and GA versions of the Text to Speech API. The GA release requires that users upgrade to the new version of the service.

A new programming model supports direct interaction between a client and the service. By using this model, a client can obtain an authentication token for communicating directly with the service. By using the token, the client can bypass the need for a server-side proxy application in IBM Cloud to call the service on its behalf. Tokens are the preferred means for clients to interact with the service.

The service continues to support the old programming model that relied on a server-side proxy to relay communications and data between the client and the service. But the new model is more efficient and provides higher throughput.

You can now pass Speech Synthesis Markup Language (SSML) to the HTTP GET and POST versions of the /v1/synthesize method. SSML is an XML-based markup language that is designed to provide annotations of text for speech synthesis applications such as the Text to Speech service. For more information about passing SSML input to the service, see Specifying input text.

The service initially supports the use of SSML only for the UK and US English, French, and German languages. The service does not support SSML for use with Italian and Spanish. When you use SSML, make sure that you do not select a voice for the audio in one of the unsupported languages. Results in this case are not meaningful.

The voices that are supported for synthesized speech that is changed and expanded. The service now supports several other voices, languages, and dialects with the /v1/synthesize methods. For more information about supported voices, see Languages and voices.

The three voices that were available at beta are renamed for GA:

• VoiceEnUsMichael is now en-US_MichaelVoice
• VoiceEnUsLisa is now en-US_LisaVoice
• VoiceEsEsEnrique is now es-ES_EnriqueVoice

The previous names of the voices continue to work with the beta version of the service (via -beta API endpoints) while that version remains available. However, you must use the new names with the GA version of the service.

You can now request the service to return audio in the Free Lossless Audio Codec (FLAC) format. The service can still return audio in the Ogg format with the Opus codec (the default) and in the Waveform Audio File Format (WAV). For more information about using audio formats with the /v1/synthesize methods, see Using audio formats.

The text that you send to the /v1/synthesize method in the URL of an HTTP GET request or in the body of an HTTP POST request is now limited to a maximum of 5 KB. The text had a maximum size of 4 MB for the beta version.

The /v1/synthesize methods now include the header X-WDC-PL-OPT-OUT to control whether the service uses the text and audio results from an operation to improve future results. Specify a value of 1 for the header to prevent the service from using text and audio results. The parameter applies only to the current request. The new header replaces the X-logging header from the beta methods. For more information, see Controlling request logging for Watson services.

For the /v1/synthesize methods, the following error codes changed:

• Error code 406 ("Not acceptable. Unsupported MIME type.") is removed.
• Error code 415 ("Unsupported Media Type") is added.
• Error code 503 ("Service Unavailable") is added.

For the GET /v1/voices method, the following error codes changed:

• Error code 406 ("Not acceptable. Unsupported MIME type.") is removed.
• Error code 415 ("Unsupported Media Type") is added.