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

# Text-to-speech

> Spoken playback of answers

The **text-to-speech** model turns written answers into spoken audio. [Observatory](/products/observatory) uses it when a member plays a reply aloud. It does not transcribe. Speech-to-text is a separate model. Duplex voice is [speech-to-speech](/products/models-speech-to-speech).

## Parameters

The production synthesizer is a low-latency neural text-to-speech model.

| Parameter | Value                                                           |
| --------- | --------------------------------------------------------------- |
| Kind      | Neural text-to-speech                                           |
| Input     | Answer text, up to 2k tokens                                    |
| Output    | Spoken audio as `mp3`                                           |
| Voice     | A selected speaking voice. Customers do not pick the checkpoint |

## Configuration

| Setting       | How it is configured                               |
| ------------- | -------------------------------------------------- |
| Where it runs | Spoken playback of Explore replies in Observatory. |
| Format        | Streaming `mp3`.                                   |
| Serving       | AiQL selects the checkpoint and the default voice. |