Large speech languages and models

Supported values

The table lists the supported values that can be used with the parameter_set request parameter.

Combining parameter_set with other parameters

When you specify parameter_set=enhanced, the preset values take precedence over any values you set individually for the parameters that the preset tunes. To set the parameters individually, omit parameter_set (or set parameter_set=default) and configure them directly.

The enhanced preset tunes the following parameters:

• sad_module
• speech_detector_sensitivity
• background_audio_suppression
• character_insertion_bias
• end_of_phrase_silence_time

Use parameter_set=enhanced to apply recommendations for the specified language in a single setting rather than configuring each parameter individually. For more information on recommended parameter configurations and general parameter tuning guidance, see Recommended parameter configurations and Recommended query parameters for large speech models.

Availability and usage

The table lists the availability based on model and Websocket usage.

{
  "action": "start",
  "content-type": "audio/wav",
  "parameter_set": "enhanced"
}

Notes

Consider the following when you use parameter_set request parameter:

• Only large speech models support parameter_set. Sending it to a non-LSM model is silently ignored.
• Five base models cover thirteen locale codes through Watson language alias chain:
  • English: en-AU, en-GB, en-IN
  • Spanish: es-AR, es-CL, es-CO, es-ES, es-MX, es-PE
  • French: fr-CA, fr-FR
  • Portuguese: pt-BR, pt-PT
  • German: de-DE

Recommended parameter configurations

The following configurations provide recommended baseline settings across supported languages.

• English (en-*)

{
  "sad_module": 2,
  "speech_detector_sensitivity": 0.9,
  "background_audio_suppression": 0.1,
  "character_insertion_bias": -1.0,
  "end_of_phrase_silence_time": 1.0
}

• French (fr-FR)

{
  "sad_module": 2,
  "speech_detector_sensitivity": 0.8,
  "background_audio_suppression": 0.0,
  "character_insertion_bias": 0.0,
  "end_of_phrase_silence_time": 1.0
}

• Spanish (es-*)

{
  "sad_module": 1,
  "speech_detector_sensitivity": 0.2,
  "background_audio_suppression": 0.0,
  "character_insertion_bias": 0.0,
  "end_of_phrase_silence_time": 1.0
}

• Brazilian Portuguese (pt-BR)

{
  "sad_module": 2,
  "speech_detector_sensitivity": 0.8,
  "background_audio_suppression": 0.0,
  "character_insertion_bias": -0.3,
  "end_of_phrase_silence_time": 1.0
}

• German (de-DE)

{
  "sad_module": 1,
  "speech_detector_sensitivity": 0.5,
  "background_audio_suppression": 0.0,
  "character_insertion_bias": -0.2,
  "end_of_phrase_silence_time": 1.0
}

Parameter guidelines

The following guidelines provide recommended parameter values and tuning strategies to help you achieve better speech recognition accuracy, reduce timeout issues, and improve overall model performance.

1. sad_module

   • Recommended: 2 (default is 1)
   • Provides improved speech/silence segmentation compared to the default CNN-based SAD
   • Helps reduce timeout issues caused by misclassification of noise as speech

2. speech_detector_sensitivity

   • Higher values lead to more aggressive speech detection
   • Large speech models generally benefit from higher sensitivity
   • Recommended range: 0.6 to 0.8 depending on language and noise conditions

3. end_of_phrase_silence_time

   • Recommended: 0.7 to 0.8 (default is 0.0)
   • Improves segmentation for conversational applications
   • Particularly beneficial for short structured utterances

4. background_audio_suppression

   • Helps improve recognition accuracy for speech in background noise
   • Should be tuned conservatively to avoid suppressing speech energy
   • Example: 0.1 for moderate noise environments

5. character_insertion_bias

   • Default: -0.22
   • Slight increase (for example, -0.1 to 0.0) can help reduce deletion errors
   • Use caution: Positive values may introduce insertion errors

Target use cases

The following recommendations are designed for conversational and interactive applications, including:

• Virtual assistants
• Healthcare voice interfaces
• Banking and insurance systems
• Customer support and contact center automation

These use cases typically involve short to medium-length utterances, frequent turn-taking interactions, and variable background noise conditions.

Supported parameters

The following parameters are commonly tuned with large speech models:

• sad_module - Selects the Speech Activity Detection (SAD) module.
• speech_detector_sensitivity - Controls how aggressively speech is detected. Higher values increase sensitivity to speech onset.
• background_audio_suppression - Controls noise suppression.
• character_insertion_bias - Controls the balance between insertion and deletion errors.
• end_of_phrase_silence_time - Defines the duration of silence (in seconds) required to finalize an utterance.

For more information about these parameters, see the Parameter summary.