TEN Logo

Sarvam just shipped an India-native text-to-speech experience: 11 Indian languages with authentic accents, a low-latency API built for real-time calls, and pronunciation tuned for real-world, domain-heavy content (all straight from their official TTS page). New sign-ups even get ₹1000 in free credits to try it.

Good news: TEN already includes a drop-in Sarvam HTTP TTS extension at ai_agents/agents/ten_packages/extension/sarvam_http_tts. You can wire it into any TEN agent in a couple of minutes.

Why Sarvam + TEN

India-first voices: one memorable voice across 11 Indian languages with authentic accents.
Real-time ready: low-latency API for live assistants and IVRs.
Domain-aware: trained on real-world Indian data to handle names, abbreviations, and numerics cleanly.
Friendly onboarding: free ₹1000 credits, with custom pricing once you scale.

What ships in TEN

Extension: ai_agents/agents/ten_packages/extension/sarvam_http_tts (manifest + tests).
Voices & controls: BCP-47 language codes, speakers (anushka, manisha, vidya, arya, abhilash, karun, hitesh), pitch, pace, loudness, and speech_sample_rate (8000/16000/22050/24000).
Model: Bulbul (model: "bulbul:v2") with base64 WAV decoding to PCM, built-in retries, optional dumping for debugging.

Quick start

1) Point your app at the extension

Add the Sarvam extension path to your TEN manifest (pattern used in ai_agents/agents/examples/voice-assistant/tenapp/manifest.json):

{
  "extensions": [
    { "path": "../../../ten_packages/extension/sarvam_http_tts" }
  ]
}

2) Configure TTS in your app property

Swap your TTS block to Sarvam (or add a new one) and set your API key (SARVAM_TTS_KEY):

{
  "type": "extension",
  "name": "tts",
  "addon": "sarvam_http_tts",
  "extension_group": "tts",
  "property": {
    "dump": false,
    "dump_path": "./",
    "params": {
      "api_subscription_key": "${env:SARVAM_TTS_KEY}",
      "target_language_code": "en-IN",
      "speaker": "anushka",
      "model": "bulbul:v2",
      "pitch": 0.0,
      "pace": 1.0,
      "loudness": 1.0,
      "speech_sample_rate": 22050,
      "enable_preprocessing": false
    }
  }
}

Other knobs you can set: output_audio_codec, endpoint (defaults to https://api.sarvam.ai/text-to-speech), and enable_preprocessing to normalize English words/numbers before synthesis.

3) Run your agent

Restart your TEN agent (or the sample voice assistant) and the TTS interface will stream Sarvam audio back to your transport. If you want debug dumps, set dump: true and inspect the PCM at dump_path.

Tips for great output

Use BCP-47 language codes (hi-IN, bn-IN, en-IN, etc.) to match Sarvam's language routing.
Keep inputs under ~1500 characters and add commas in long numbers for clean pronunciation (10,000 instead of 10000).
Tune pitch, pace, and loudness per voice; Bulbul responds naturally to moderate adjustments.
For domain-heavy text, set enable_preprocessing: true and supply domain-specific terms directly—Sarvam is trained to keep them intact.

Where to learn more

Sarvam TTS docs: https://docs.sarvam.ai/api-reference-docs/text-to-speech-rest/convert
TEN extension source: ai_agents/agents/ten_packages/extension/sarvam_http_tts
Example manifest wiring: ai_agents/agents/examples/voice-assistant/tenapp/manifest.json

India-first voices with low latency and a generous trial: a perfect fit for TEN agents that need to sound local, fast, and trustworthy. Give Sarvam a spin and ship a truly multilingual voice experience today.

Sarvam TTS in TEN: India-first voices in minutes