> ## 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.

# Transcriber

> Providers that turn what the caller says into text

The transcriber converts the caller's speech to text, which is what the language model reads to decide its reply. It also produces the transcript on the run record.

Set your default transcriber on [Model Configurations](/talk/configurations/inference-providers), and override it per workflow where you need a different one. Credentials come from a [connection](/talk/configurations/connections).

## Supported providers

| Provider                                                    | Notes                                                   |
| ----------------------------------------------------------- | ------------------------------------------------------- |
| [Deepgram](/talk/configurations/providers/deepgram)         | Default transcriber. Also covers voice                  |
| [AssemblyAI](/talk/configurations/providers/assemblyai)     |                                                         |
| Speechmatics                                                |                                                         |
| Gladia                                                      |                                                         |
| [ElevenLabs](/talk/configurations/providers/elevenlabs)     | Also covers voice                                       |
| [Cartesia](/talk/configurations/providers/cartesia)         | Also covers voice                                       |
| [OpenAI](/talk/configurations/providers/openai)             | Also covers language models and voice                   |
| [Azure Speech](/talk/configurations/providers/azure-speech) | Needs a region. Also covers voice                       |
| [Google Cloud](/talk/configurations/providers/google-cloud) | Also covers voice                                       |
| [Groq](/talk/configurations/providers/groq)                 | Also covers language models                             |
| [Sarvam](/talk/configurations/providers/sarvam)             | Indian languages. Also covers language models and voice |

<Note>
  A provider listed here as covering more than one service needs only **one** connection.
</Note>

## Choosing a language

Set the language to the one your callers actually speak. Transcription accuracy drops sharply when the language is wrong or left on a default that does not match, and every downstream step reads the transcript — so a bad transcription becomes a bad reply.

Language support varies by provider and by model. Check the provider's own documentation for which languages a given model handles before committing to it.

## Accuracy matters more than it looks

If the agent keeps misunderstanding callers, check the transcript before changing the prompt. The model can only respond to what it was given, and a transcription error looks identical to a reasoning error from the outside.

The [Tracing](/talk/configurations/tracing) view shows the transcription for each turn alongside what the model did with it, which is the fastest way to tell the two apart.