Voice & language catalog

A session picks its speech-to-text model, language, and text-to-speech voice either at connect time (the client’s pipeline config) or mid-call (the brain’s configure_stt / configure_tts). This page is the catalog of allowed values.

Deployment-specific

The concrete set of enabled models and installed voices is a property of the deployment (each engine is opt-in server-side). The values below are what ships in the public demos and are the stable, documented surface.

Speech-to-text

Model (stt.model)

Value	What it is
vql-stt (default)	A composite router covering English plus 22 Indic languages. It picks the underlying engine from language — you never name an engine directly. No language → English. Use this.
nemotron-streaming	A directly-pinned engine, for comparison/testing.
cohere-transcribe	A directly-pinned engine, for comparison/testing.

For almost every agent, leave the model as vql-stt and just set the language.

Language (stt.language)

vql-stt supports English (en) plus these 22 Indic languages:

Code	Language	Code	Language	Code	Language
as	Assamese	bn	Bengali	brx	Bodo
doi	Dogri	kok	Konkani	gu	Gujarati
hi	Hindi	kn	Kannada	ks	Kashmiri
mai	Maithili	ml	Malayalam	mr	Marathi
mni	Manipuri	ne	Nepali	or	Odia
pa	Punjabi	sa	Sanskrit	sat	Santali
sd	Sindhi	ta	Tamil	te	Telugu
ur	Urdu

A mid-call language change (configure_stt(language_hint=...)) that crosses the English↔Indic boundary is applied at the next end-of-turn, never mid-utterance.

STT session knobs

Passed via configure_stt (applies live). Defaults shown:

Knob	Purpose
language_hint	Switch STT language (any code above); crosses the English↔Indic boundary at the next end-of-turn
VAD knobs	vad_confidence, vad_min_volume, vad_start_frames, vad_stop_frames_to_trigger_update, vad_eager_frames, vad_barge_in_ms, resume_frames, min_segment_speech_frames, confidence_tail_ms — voice-activity / end-of-turn tuning

These are exactly the arguments Session.configure_stt(...) accepts.

Text-to-speech

Voice (tts.voice)

Voice IDs are engine/voice — the engine prefix is required (a bare name is rejected; there is no implicit default). The prefix, not any model field, selects the engine.

The engine the public demos use is omnivoice, a voice-cloning model with two personas:

Voice ID	Persona
omnivoice/gauri	Female
omnivoice/gaurav	Male
omnivoice/auto	Model-chosen voice (no fixed persona)

Both personas default to Hindi and have dedicated reference clips for hi, en, bn, gu, kn, ml, mr, pa, ta, te; other languages fall back to the Hindi clip.

Note

Other TTS engines (supertonic, indicf5, kokoro) exist and may be enabled in a given deployment, with their own voice IDs. The omnivoice/* voices above are the shipping catalog. tts.model is not a real knob — the engine is chosen entirely by the voice-id prefix.

Setting it

• At connect (client): the React SDK’s pipeline prop, e.g.

  pipeline: {
    stt: { model: "vql-stt", language: "en" },
    tts: { voice: "omnivoice/gauri", language: "en" },
  }

• Mid-call (brain): session.configure_stt(language_hint="hi") and session.configure_tts(voice="omnivoice/gaurav"). STT applies live; TTS applies at the next inference. See Handling a conversation.

The lead_qual demo switches both STT and TTS language mid-call for multilingual qualification — a good worked example.