# Sarvam STT (Saaras and Saarika)

> Use Sarvam speech-to-text for Indian-language voice agents, including the codemix mode for Hinglish and other mixed-script speech.

Source: https://edesy.in/docs/voice-agent/stt/sarvam

Sarvam is a speech-to-text provider built for Indian languages. It runs on the default Edesy runtime. Choose it when callers speak Hindi, Tamil, Telugu, Bengali or another Indian language, or mix English into an Indian language. Set the `codemix` mode so English words stay in Latin script instead of being turned into the native script.

## At a glance

| Item | Value |
|------|-------|
| Provider ID | `sarvam` (the dashboard also stores `sarvam_stt`) |
| Runtime | Default runtime |
| Default model | `saaras:v3` |
| Other models | `saarika:v2.5`, `saarika:v2` |
| Languages | English (India), Hindi, Tamil, Telugu, Kannada, Malayalam, Bengali, Gujarati, Marathi, Punjabi, Odia |
| Mixed speech | `mode: "codemix"` keeps English words in Latin script |
| Your own key (BYOK) | Supported. Add a Sarvam key in **Settings → API Keys** |

## When to use Sarvam

- Your callers speak an Indian language and often switch to English mid-sentence ("mera order kab deliver hoga?").
- You want transcripts that your prompt, tools and extraction can read reliably, with English product names and numbers kept as they were said.
- You are building in Malayalam, Tamil or another language where code-mixed speech is common.

For English-only agents, [Deepgram](https://edesy.in/docs/voice-agent/stt/deepgram) is usually the simpler default.

## Set it up

1. Open your agent and go to the **Voice** settings.
2. Under **Speech-to-text**, choose **Sarvam AI**.
3. Pick the model. Keep `saaras:v3` unless you have a reason to change it.
4. Set the agent language (for example Hindi or Hinglish).
5. Place a test call from the [Test Center](https://edesy.in/docs/voice-agent/agent-setup/evals) and check the transcript.

Through the API, set the provider on the agent:

```json
{
  "language": "hindi_english",
  "sttProvider": "sarvam",
  "sttConfig": {
    "model": "saaras:v3",
    "mode": "codemix"
  }
}
```

## Settings

| Setting | Values | What it does |
|---------|--------|--------------|
| `model` | `saaras:v3`, `saarika:v2.5`, `saarika:v2` | The Sarvam model. `saaras:v3` is the default |
| `mode` | `codemix` (recommended for mixed speech) | Keeps English words in Latin script inside Indian-language transcripts |

## FAQs

**Is Sarvam STT included in the per-minute price?**
Sarvam STT has no add-on charge in the dashboard catalog. See [provider costs](https://edesy.in/docs/voice-agent/billing/provider-costs) for the current list.

**Should I use `codemix` for pure Hindi calls?**
It does no harm. It matters most when callers use English words, which is common on support and sales calls.

**Can I use Sarvam for TTS too?**
Yes. See [Sarvam Bulbul TTS](https://edesy.in/docs/voice-agent/tts/sarvam).

**Can I bring my own Sarvam key?**
Yes. Add it under **Settings → API Keys**. See [BYOK and key pools](https://edesy.in/docs/voice-agent/workspace/byok).

**Which Indian languages does Edesy support overall?**
Edesy supports 75+ languages, including 14 Indian languages and Hinglish and bilingual modes. See [language support](https://edesy.in/docs/voice-agent/languages).

## Related

- [STT providers overview](https://edesy.in/docs/voice-agent/stt)
- [Soniox STT](https://edesy.in/docs/voice-agent/stt/soniox)
- [Language selection](https://edesy.in/docs/voice-agent/agent-setup/language-selection)