Speech-to-Text (STT)

The STT layer transcribes participant speech in real time using automatic speech recognition (ASR). You can select a model optimized for your use case and language requirements.

STT models

Select an STT model using the stt_engine parameter in the layers.stt object. The following models are available:

Model	Description
`tavus-auto` (default)	Automatically selects the best STT model for the conversation’s language. Recommended for most use cases.
`tavus-parakeet`	Highest throughput, lowest latency for English and European languages.
`tavus-soniox`	Purpose-built for Indian languages with broad multilingual coverage.
`tavus-whisper`	Broad multilingual coverage across all supported languages.
`tavus-deepgram-medical`	Domain-specific English STT optimized for clinical and healthcare vocabulary. English only.
`tavus-advanced`	Deprecated. Still active but not recommended for new integrations.

tavus-auto is the default. Use it unless you have a specific language or domain requirement. It automatically routes to the best model for each conversation.

Choosing the right model

A language is listed for a model only if both STT and TTS coverage are available.

Category	Recommended model	Supported languages
General purpose	`tavus-auto`	All supported languages
Indic languages	`tavus-soniox`	Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Punjabi, Tamil, Telugu + broad support for all other languages
English + European	`tavus-parakeet`	Bulgarian, Croatian, Czech, Danish, Dutch, English, Finnish, French, German, Greek, Hungarian, Italian, Polish, Portuguese, Romanian, Russian, Slovak, Spanish, Swedish, Ukrainian
Broad multilingual	`tavus-whisper` or `tavus-soniox`	All supported languages
Medical (English)	`tavus-deepgram-medical`	English

Smart Language Detection requires either tavus-auto, tavus-soniox, or tavus-whisper. Name the languages it detects with languages.

Which languages get heard

languages is what decides that. Set it on the PAL, or on a single conversation, and those are the languages STT listens for:

{ "languages": ["en", "es", "pt"] }

With tavus-auto, Tavus then picks a model that covers your whole set, so you do not have to match engines to languages yourself. With languages set on neither the PAL nor the conversation, the conversation runs in English, the same as sending ["en"]. See Language support for the full list and the accepted values.

Configuring the STT layer

Define the STT layer under the layers.stt object.

`stt_engine`

Set the STT model for transcription:

"stt": {
  "stt_engine": "tavus-auto"
}

`hotwords`

Use this to prioritize certain names or terms that are difficult to transcribe.

"hotwords": "Roey is the name of the person you're speaking with."

The above helps the model transcribe “Roey” correctly instead of “Rowie.”

Use hotwords for proper nouns, brand names, or domain-specific language that standard STT engines might struggle with.

Example configuration

Below is an example PAL with a configured STT layer using the recommended tavus-auto engine:

{
  "pal_name": "Customer Service Agent",
  "system_prompt": "You assist users by listening carefully and providing helpful answers.",
  "pipeline_mode": "full",
  "default_face_id": "r90bbd427f71",
  "layers": {
    "stt": {
      "stt_engine": "tavus-auto",
      "hotwords": "Roey is the name of the person you're speaking with."
    }
  }
}

Refer to the Create PAL API for a complete list of supported fields.

Getting Started

Guides

Build

Deploy

Debug

Resources

Speech-to-Text (STT)

STT models

Choosing the right model

Which languages get heard

Configuring the STT layer

`stt_engine`

`hotwords`

Example configuration

​STT models

​Choosing the right model

​Which languages get heard

​Configuring the STT layer

​stt_engine

​hotwords

​Example configuration

STT models

Choosing the right model

Which languages get heard

Configuring the STT layer

`stt_engine`

`hotwords`

Example configuration