The sarvam engine on Wixzel Voice: Sarvam, Indian languages end to end
Indian languages end to end. Sarvam Saaras v3 transcribes, the Sarvam-105B model reasons, and Bulbul v3 speaks, so Hindi, Tamil, Telugu and the other Indian languages Sarvam supports are handled by models built for them rather than by an English pipeline with a language flag.
Last updated
What does each model in the sarvam engine cost on Wixzel Voice?
A composed engine is three stages, each billed in its own unit. The per-minute column is what that unit adds up to at typical speech rates. The pipeline label is Sarvam, Indian languages end to end.
| Stage | Model | Unit price | Per minute |
|---|---|---|---|
| Speech-to-text | sarvam/saaras-v4 | $0.0001 per second | $0.0076 |
| Speech-to-text | sarvam/saaras-v3-realtime | $0.0001 per second | $0.0076 |
| Language model | sarvam/sarvam-105b | $0.0002 per 1k tokens in | $0.0005 |
| Language model | sarvam/sarvam-105b | $0.0009 per 1k tokens out | $0.0004 |
| Text-to-speech | sarvam/bulbul-v3 | $0.0457 per 1k characters | $0.0411 |
| Text-to-speech | sarvam/bulbul-v4-flash | $0.0152 per 1k characters | $0.0137 |
| Orchestration | wixzel/orchestration | $0.0120 per minute | $0.0120 |
Rounded up, never down. Where two models are listed for one stage the engine picks by language; the engine price above assumes the dearer path. All prices on the model price list.
Frequently asked questions
What is the sarvam voice engine?
How much does the sarvam engine cost per minute?
Which models does the sarvam engine use?
sarvam/saaras-v4, sarvam/saaras-v3-realtime, sarvam/sarvam-105b, sarvam/bulbul-v3, sarvam/bulbul-v4-flash. Each stage is billed in its own unit: seconds of speech recognised, tokens in and out of the model, and characters synthesised.Do I need my own sarvam account to use sarvam?
wv_live_… API key and a prepaid balance cover the sarvam engine and every other engine; there are no provider keys to bring.How do I use the sarvam engine in a call?
POST /v1/agents, for example the stt, llm and tts refs listed above, then place a call with POST /v1/calls. GET /v1/engines lists what is available right now; an engine whose provider is degraded disappears from that list before calls start failing.Is the sarvam price per minute a flat rate?
GET /v1/usage/events. A quieter or shorter turn costs less; a long monologue costs more.What do developers build with the sarvam engine?
The stt provider selects the pipeline, and Sarvam's models are trained on Indian languages and accents rather than adapted from English. It is the engine to reach for whenever the caller's first language is not English.
Many clients' callers speak Hindi, Tamil, Malayalam or another Indian language, and sarvam is built for 11 Indian languages end to end. Where cost matters most, gemini-live is $0.0491 per minute and speaks 97 languages, Hindi, Tamil and Malayalam among them. Each client's agent can use a different engine.
Which guides build on the sarvam engine?
Which other voice engines does Wixzel Voice run?
Deepgram + GPT-6 Luna + ElevenLabs
Gemini Live, native audio
Deepgram Voice Agent
OpenAI GPT-Live + GPT-6 Luna
OpenAI Realtime, gpt-realtime-2.1-mini
xAI Grok Voice
Name a model for each stage and see the minute it adds up to.
Build your first voice agent