Transcriber tab

Fields

All fields live under transcriber in the agent object. Some fields apply to one provider only.

Field	Type	Default	Allowed	Provider	What it does
provider	string	cartesia	cartesia, deepgram, sarvam	all	The speech-to-text service.
model	string	ink-2	ink-2 (Cartesia), nova-3 (Deepgram), saaras:v3-realtime or saaras:v4 (Sarvam)	all	The model. If it doesn’t match the provider, Vaakyo replaces it: ink-2 for Cartesia, nova-3 for Deepgram, saaras:v3-realtime for Sarvam.
turn_detection	string	balanced	responsive, balanced, patient	Cartesia, Sarvam	How quickly the agent answers once the caller pauses. On Sarvam: a pause of 0.5 s, 0.7 s or 1.1 s.
language	string	hi	Deepgram: hi, multi, en-IN, en. Sarvam: auto, hi-IN, en-IN, bn-IN, gu-IN, kn-IN, ml-IN, mr-IN, od-IN, pa-IN, ta-IN, te-IN	Deepgram, Sarvam	The language the caller speaks. auto makes Sarvam detect it. For Sarvam, hi, en and multi are accepted and saved as hi-IN, en-IN and auto. Cartesia ignores it and detects the language itself.
endpointing_ms	integer	300	10 to 3000	Deepgram	Milliseconds of silence that end the caller’s turn.
keyterms	array of strings	[]	up to 100	all	Words the recogniser should expect, such as doctor or product names. Cartesia uses up to 100, Deepgram the first 50, Sarvam the first 50 on saaras:v4 only.
fallbacks	array	[]	up to 2	all	Backup transcribers, connected in order when the transcriber fails. See Fallback transcribers.

Cartesia example

{
  "transcriber": {
    "provider": "cartesia",
    "model": "ink-2",
    "turn_detection": "patient",
    "keyterms": ["Dr. Kulkarni", "Aundh", "physiotherapy"]
  }
}

Deepgram example

{
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-3",
    "language": "multi",
    "endpointing_ms": 400,
    "keyterms": ["Dr. Kulkarni"]
  }
}

Sarvam example

{
  "transcriber": {
    "provider": "sarvam",
    "model": "saaras:v3-realtime",
    "language": "hi-IN",
    "turn_detection": "balanced"
  }
}

Choosing a provider

	Cartesia ink-2	Deepgram nova-3	Sarvam saaras
Language	Detects it on its own (en, hi, fr, ja, es)	You set it: hi, multi (switches between languages), en-IN, en	You set it (11 Indian languages and Indian English), or auto to detect it
End of turn	A turn-detection model, tuned with turn_detection	A fixed silence, endpointing_ms	Sarvam’s voice-activity detection: a pause set by turn_detection
Good for	Callers who mix Hindi and English; natural pauses	A known single language; strict control of the pause length	Indian languages beyond Hindi; Hindi agents that also speak with a Sarvam voice

Sarvam writes Hindi in Devanagari, including English words said in Hindi (“अपॉइंटमेंट”). The LLM reads it fine, and transcripts show it as written.

Your workspace may not be allowed to use every provider. GET /api/v1/providers shows which providers your workspace can use; saving an agent with a provider it can’t use returns 422.

Turn detection

The agent starts its reply when the caller’s turn ends, so this setting trades speed against cutting the caller off.

turn_detection	Behaviour
responsive	Answers quickly. Good for short answers (“yes”, “no”, a date).
balanced	The default.
patient	Waits longer. Good for slow speakers, elderly callers, or callers who think aloud.

On Deepgram, raise endpointing_ms (for example to 500) if the agent talks over callers who pause mid-sentence, and lower it if replies feel slow.

If a caller pauses mid-sentence and the turn ends early, Vaakyo joins the two parts into one message for the LLM when the agent has not spoken in between. The transcript still shows them as two turns.

Fallback transcribers

transcriber.fallbacks lists up to two backup transcribers. When the transcriber fails to connect at the start of a call, or its connection drops during the call, the next one is connected and used for the rest of the call. Audio that arrives while it connects is kept and sent to it.

Field	Type	Default	What it does
provider	string	deepgram	cartesia, deepgram or sarvam.
model	string	nova-3	Replaced to fit the provider, as for the primary (ink-2, nova-3 or saaras:v3-realtime).
language	string	hi	Deepgram and Sarvam, as for the primary.

Backups use the primary’s keyterms, turn_detection and endpointing_ms. A backup can’t repeat the primary or an earlier backup (same provider, model and, on Deepgram and Sarvam, language); saving returns 422. Your workspace must be allowed to use the backup’s provider.

{
  "transcriber": {
    "provider": "cartesia",
    "model": "ink-2",
    "fallbacks": [{"provider": "deepgram", "model": "nova-3", "language": "multi"}]
  }
}

Each switch is a fallback.used event on the call’s timeline, and the call’s providers.stt names the transcriber used last. Without fallbacks nothing changes: a transcriber that can’t connect fails the call, and one that stops mid-call ends it with an error. See Webhook events.

When the transcriber’s provider is at its concurrency limit as a call starts, the first fallback transcriber with room takes the call (a capacity.fallback event). See Webhook events.

Usage

The seconds of caller audio sent to the transcriber are counted in the call’s usage.stt_seconds.