Multilingual Hindi Voice Agent Setup: Sarvam AI & n8n

Why do traditional voice bots fail Indian SMB callers?

Everyone believes global voice models handle any language now. Here’s why that’s wrong.

Off-the-shelf Western engines fail on conversational Indian English, regional dialects, and code-mixed Hinglish. A customer says, “Mera order deliver kab hoga? Tracking link check kar do.” Standard speech-to-text engines drop crucial keywords. They hallucinate weird English approximations instead.

Western voice stacks also break your budget. Dollar-denominated models for transcription and voice generation inflate the cost of every call. That makes automated phone operations impossible for high-volume support desks, logistics dispatchers, or lead hotlines.

Connecting specialized Indian language models to workflow automation solves this gap. If you’ve already built telephony routing using our guide to Build an Inbound Exotel AI Voice Agent with n8n, you can easily layer Sarvam AI into that pipeline. You’ll get native speech recognition and synthesis tuned specifically for Indian phone networks.

Caller Audio → n8n Webhook → Sarvam Saaras STT → LLM Processing → Sarvam Bulbul TTS → Audio Response

Step 1: Provision your Sarvam AI API credentials

Create an account on the Sarvam AI dashboard. According to Sarvam AI Speech-to-Text cost: ₹30/hour base rate, new Sarvam AI accounts receive ₹100 in universal non-expiring free credits on signup, which covers roughly two hours of base speech-to-text processing for testing and prototyping.

After signing up:

1. Go to the API Keys section of the dashboard.
2. Generate an API Key. Store it safely in your password manager.
3. Track your usage in the console. According to Sarvam AI Speech-to-Text cost: ₹30/hour base rate, Sarvam AI charges a base rate of ₹30 per hour for Speech-to-Text across real-time, streaming, and batch modes, with billing calculated per second of audio.

Step 2: Ingest audio via an n8n Webhook node

Open your self-hosted or cloud n8n instance. Create a new workflow.

1. Add a Webhook node.
2. Set the HTTP Method to POST.
3. Set the binary property to receive audio files (like data or file).
4. Point your telephony provider or WhatsApp media listener callback to this webhook URL. You can use Exotel or Twilio here.

According to Build workflows with Sarvam AI in n8n, Sarvam’s n8n community integration enables voice workflows to mirror patterns used in LiveKit and Pipecat by connecting audio inputs and outputs directly across standard n8n nodes.

Step 3: Transcribe audio using Sarvam Saaras STT

Configure an HTTP Request node to forward the audio file to the Saaras Speech-to-Text endpoint.

According to REST | Sarvam API Docs, Sarvam’s Saaras v3 and v4 Speech-to-Text APIs accept multipart audio files in formats like WAV, MP3, AAC, and PCM, supporting specialized Indic modes including transcribe, verbatim, translit, and codemix.

Configure the node parameters like this:

• Method: POST
• URL: https://api.sarvam.ai/speech-to-text
• Authentication: Header Auth (api-subscription-key: YOUR_SARVAM_API_KEY)
• Body Content Type: Multipart Form-Data
• Form Fields:
  • file: Pick binary data from the incoming webhook payload.
  • model: Set to saaras:v3 or saaras:v4.
  • language_code: hi-IN (or leave empty for language auto-detection).
  • mode: Set to codemix to capture Hindi and English phrasing seamlessly.

Step 4: Run conversational logic via LLM

Pass the transcript returned by Saaras into an LLM node or an AI Agent node in n8n.

1. Map {{ $json.transcript }} into the user prompt field.
2. Write system instructions that enforce short, conversational responses. Tell the model to respond in plain Hindi using Devanagari script or conversational Hinglish. Tell it to skip markdown formatting, bullet points, and tables.
3. Keep the token limit concise, ideally under 100 words. This ensures your voice generation step finishes quickly with minimal latency.

Step 5: Synthesize spoken Hindi using Sarvam Bulbul v3 TTS

Add another HTTP Request node to turn the model’s text answer back into natural audio with Sarvam’s Bulbul engine.

According to REST | Sarvam API Docs, Sarvam’s Bulbul v3 Text-to-Speech API supports up to 2,500 characters per request, providing over 30 distinct voices across 11 Indian languages, returning base64-encoded audio. According to Text to Speech & AI Voice Generator for Indian Languages – Sarvam AI, Bulbul v3 incorporates text normalization that automatically expands Indian abbreviations, addresses, PIN codes, and currency amounts in rupees into proper colloquial spoken Hindi.

Set up the request body as JSON:

{
  "inputs": ["{{ $json.response_text }}"],
  "target_language_code": "hi-IN",
  "speaker": "shubh",
  "pitch": 0,
  "pace": 1.0,
  "loudness": 1.5,
  "speech_sample_rate": 8000,
  "enable_preprocessing": true,
  "model": "bulbul:v3"
}

Step 6: Dispatch base64 audio response to the caller

1. The Bulbul v3 API returns a JSON object with an array of base64-encoded audio strings.
2. Add a final Respond to Webhook node in n8n.
3. Pass the base64 string directly, or convert the binary audio buffer depending on what your telephony provider expects.
4. Close the session, or loop the output back into your listening hook for multi-turn conversations.

If your business needs production-grade Indic AI calling workflows, custom telephony bridges, or automated customer support pipelines, LoomVector designs and implements tailored voice infrastructure for growing companies.