> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sentivue.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Sarvam

> Connect Sarvam to SentiVue Talk

Sarvam builds models for Indian languages — 22 Indic languages plus Indian-accented English, across text, speech recognition and speech synthesis. In SentiVue Talk it drives the **language model**, **voice** and **transcription**. One Sarvam connection can run an entire pipeline, with no second provider needed.

## What you need

| Field   | Required |
| ------- | -------- |
| API key | Yes      |

One key covers all three services.

## Getting your API key

1. Sign in at [dashboard.sarvam.ai](https://dashboard.sarvam.ai/admin).
2. Go to the **API Keys** section and create a new key.
3. Copy it — the dashboard shows the value once and it cannot be retrieved after you navigate away. Store it somewhere safe.

<Warning>
  Sarvam runs on prepaid credits. When the balance reaches zero every request fails, which mid-call means the agent stops responding. Credits themselves do not expire, but a balance that quietly drains is a common cause of an agent that worked all month and stops. Track the balance on the dashboard's usage page.

  Rate limits apply to the **account**, not the key — every key you create shares one pool, so adding a key does not add headroom.
</Warning>

## Adding the connection

1. Go to **Connections** and choose **Sarvam**.
2. Paste the API key.
3. Choose which services this connection drives — language model, voice, transcription, or any combination — and save.

The key is not checked against Sarvam when you save, so a wrong or expired key first shows up as a failure on a live call. Place a test call after connecting to confirm it works.

## What you can select afterwards

Choose your models on [Model Configurations](/talk/configurations/inference-providers).

Languages are BCP-47 codes with an Indian region — `hi-IN`, `ta-IN`, `en-IN`. Coverage differs between the three services, so check each part of the pipeline you put on Sarvam. Sarvam's [models page](https://docs.sarvam.ai/api/getting-started/models) lists the current models and their coverage.

**Language model** — prefer a conversational variant over a reasoning-oriented one.

**Voice** — pick a **speaker** by name alongside the language. The list differs between model versions. Speakers are on Sarvam's [text-to-speech guide](https://docs.sarvam.ai/api/api-guides-tutorials/text-to-speech/rest-api).

**Transcription** — set the language rather than relying on auto-detection.

## Using Sarvam for only some services

The connection does not have to drive all three roles. Turn off the ones you do not want, and SentiVue Talk uses your other connections for them — Sarvam for voice and transcription with the language model on [OpenAI](/talk/configurations/providers/openai), or Sarvam for the language model with an English-only [Deepgram](/talk/configurations/providers/deepgram) transcriber. Change this at any time on the connection.

This is worth doing when only part of your call is in an Indian language. Sarvam is strongest where the caller speaks Hindi, Tamil, Telugu, Bengali or another Indic language, or code-mixes with English; for calls entirely in non-Indian languages, another [voice provider](/talk/configurations/voice) or [transcriber](/talk/configurations/transcriber) will cover more ground.

## Troubleshooting

<AccordionGroup>
  <Accordion title="The connection saved fine but calls fail immediately">
    The key is not verified at save time, so a typo, a trailing space or a rotated key only surfaces on a call. Recreate the key on the Sarvam dashboard, paste it again, and place a test call.
  </Accordion>

  <Accordion title="The agent goes silent partway through a call">
    Two causes look identical from the call. Check the credit balance first — at zero, every request fails. If credits remain, you have hit a rate limit: limits are per account and shared by all your keys, so several concurrent calls reach them far sooner than testing alone suggests. Raise your plan's limits, or move one role to another provider.
  </Accordion>

  <Accordion title="Transcripts are poor, or come back in the wrong script">
    Set the language explicitly instead of leaving detection to guess. If callers code-mix English with an Indian language, choose a model and output mode built for that — Sarvam offers transcription in the original script, romanized output, and code-mixed output, and they produce visibly different transcripts from the same audio.
  </Accordion>

  <Accordion title="A speaker or voice name is rejected">
    Speaker names are tied to a specific Bulbul version, and a name valid on one version does not carry to another. Check the name against the version you selected on Sarvam's [text-to-speech guide](https://docs.sarvam.ai/api/api-guides-tutorials/text-to-speech/rest-api).
  </Accordion>

  <Accordion title="The language I need works for transcription but not for voice">
    Coverage is not the same across the three services — transcription reaches more Indic languages than the voices do. Keep transcription on Sarvam and move the voice to another [voice provider](/talk/configurations/voice) as a separate [connection](/talk/configurations/connections).
  </Accordion>
</AccordionGroup>