Sarvam transcriber (speech to text)

​

1. What is Sarvam STT?

​

2. Key features of Sarvam STT

• Indian Language Expertise: Trained on Indian language data rather than adapted from an English-first baseline, giving high accuracy on regional accents and code-mixed speech.
• Real-Time Streaming: Built for streaming transcription over WebSocket with low latency, so conversations keep their natural rhythm.
• Code-Mixed Speech Recognition: Handles seamless switching between English and an Indian language inside a single sentence — the common case on Indian phone calls.
• Broad Language Coverage: Sarvam’s Saaras models cover 23 languages — 22 Indian languages plus English.
• Automatic Language Detection: Saaras models can detect the spoken language rather than requiring it up front.
• Global English Accents: saaras:v4 adds non-Indian English accents on top of Indian English, useful for agents that take calls from abroad.
• Keyterm Prompting: saaras:v4 takes a list of names, places, brands and technical terms to bias recognition towards.

​

3. How Bolna uses Sarvam for STT

• Real-Time Indian Language Processing: Bolna streams call audio to Sarvam’s realtime STT API and receives transcripts as the caller speaks, so the agent can respond without waiting for the utterance to finish being processed.
• Regional Language Voice Agent Support: Agents handle conversations in Hindi, Bengali, Tamil, Telugu, and the other supported languages with accuracy that generic multilingual models do not reach on Indian speech.
• Accent-Aware Transcription: Sarvam’s training on diverse Indian accents keeps transcription accurate across regions and demographics.
• Code-Switching Support: Callers who switch between English and a regional language mid-sentence are transcribed correctly, without the agent needing to reconfigure its language.
• Turn Detection Handled by Bolna: Endpointing and voice activity detection are Bolna-side settings (endpointing, vad_threshold, vad_prefix_padding_ms) that apply regardless of transcriber, so you tune them the same way you would for any other provider.

​

4. List of Sarvam models supported on Bolna AI

​

5. Supported languages

​

The other 11 Saaras languages

​

6. Configuration

{
  "transcriber": {
    "provider": "sarvam",
    "model": "saaras:v4",
    "language": "hi",
    "stream": true,
    "endpointing": 500
  }
}

​

Keyword and context biasing

"keywords": "Bolna, PhonePe, Flipkart, New Delhi",
"context": "Inbound support line for an electronics retailer in India. Callers ask about order status and returns, and often read out order numbers."

• keywords is sent as Sarvam’s keyterms, up to 50 terms of up to 64 characters each. A list over either limit is rejected when you save the agent. Multi-word phrases such as New Delhi count as one term.
• Per-term weights are not supported, so Bolna strips any term:number suffix before sending.
• context is sent as Sarvam’s prompt, a plain description of the call rather than a term list.
• Both are set when the call connects and apply for the whole call.

​

Conclusion

• Pair it with Sarvam TTS for a fully Indic voice pipeline
• Give each language its own transcriber with multilingual support
• You can also connect your own Sarvam account and use it with Bolna AI