Sarvam

Sarvam AI provides a modular speech and language platform with deep support for Indian languages and code-mixed speech. Unlike bundled voice-agent APIs, the Teler bridge orchestrates three independent Sarvam services: speech-to-text (STT), a conversational LLM, and text-to-speech (TTS), giving you full control over each stage.

Note

This bridge is Node-only; a Python variant is not yet available. Each pipeline stage is independently configurable: choose the STT model, LLM, TTS model, language, and speaker. If the primary LLM becomes unavailable, the bridge automatically falls back to a secondary model.

Reference repos

Node teler-sarvam-node-bridge

Prerequisites

• A Sarvam account with an API key
• A Teler account with an active API key
• One Teler phone number on your account
• An ngrok auth token (for local development)
• Docker and Docker Compose

Configuration

Variable	Required	Default	Description
SARVAM_API_KEY	yes	must be provided	Your Sarvam API key.
TELER_API_KEY	yes	must be provided	Your Teler API key.
NGROK_AUTHTOKEN	yes (local)	none	ngrok auth token.
SERVER_DOMAIN	auto	none	Public URL exposed to Teler. Detected from ngrok.
SARVAM_STT_MODEL	no	saaras:v3-realtime	Speech-to-text model identifier.
SARVAM_LLM_MODEL	no	sarvam-105b-conversations	Primary conversational LLM.
SARVAM_LLM_MODEL_FALLBACK	no	sarvam-105b	Secondary LLM if primary is unavailable.
SARVAM_TTS_MODEL	no	bulbul:v3	Text-to-speech engine.
SARVAM_LANGUAGE_CODE	no	en-IN	Language code (e.g. en-IN for English-India, hi-IN for Hindi).
SARVAM_SPEAKER	no	shubh	TTS voice identity.
STATIC_RESPONSE_TEXT	no	Default greeting	Greeting played when the call connects.

5-minute setup

git clone https://github.com/frejun-tech/teler-sarvam-node-bridge
cd teler-sarvam-node-bridge
cp .env.example .env
docker compose up -d --build

Initiate a call

curl -X POST https://your-domain/api/v1/calls/initiate-call \
  -H "Content-Type: application/json" \
  -d '{
    "fromNumber": "+91XXXXXXXXXX",
    "toNumber":   "+91XXXXXXXXXX"
  }'

Architecture

The bridge opens independent WebSocket connections to each Sarvam service and orchestrates the pipeline:

1. Caller audio in: raw audio frames from the Teler call are forwarded to Sarvam’s realtime STT endpoint, transcribed according to SARVAM_STT_MODEL.
2. Transcription → LLM: the transcribed text is sent to Sarvam’s chat-completions API, which generates a response using the primary LLM (SARVAM_LLM_MODEL). If that model is unavailable, the bridge falls back to SARVAM_LLM_MODEL_FALLBACK.
3. LLM response → TTS: the generated text is streamed to Sarvam’s realtime TTS endpoint, synthesized to speech using the configured SARVAM_TTS_MODEL and SARVAM_SPEAKER.
4. Agent audio out: synthesized audio is forwarded back to the Teler call as binary frames.
5. Barge-in: when the caller speaks over the agent, the bridge clears buffered audio to allow interruption.
6. Greeting: a STATIC_RESPONSE_TEXT greeting is played when the call first connects.

Troubleshooting

Symptom	Likely cause	Fix
401 from Sarvam	Invalid SARVAM_API_KEY	Rotate the key; check casing in .env
Agent misunderstands caller	Wrong language code	Set SARVAM_LANGUAGE_CODE to match the caller’s language (e.g. hi-IN for Hindi, en-IN for English-India)
No greeting on call connect	Missing or empty STATIC_RESPONSE_TEXT	Set a non-empty greeting in .env
Responses slow after first message	Primary LLM overloaded	The bridge retries the fallback LLM automatically; check Sarvam status

Next steps

Other recipes OpenAI Realtime, Gemini, ElevenLabs, Vapi, and more.

Bridge architecture The conceptual model behind every recipe.

Sarvam docs Detailed guides for STT, TTS, and LLM models.