Speech-to-speech - AiQL Handbook

The speech-to-speech model is a realtime duplex audio model. A member speaks, the model replies in speech in the same session, and it can call the same workspace tools as answers. It is the call path in Observatory, not file transcription. File audio uses speech-to-text. Playback of a written reply uses text-to-speech.

Parameters

The production realtime model accepts speech, text, and images, and returns speech and text.

Parameter	Value
Kind	Realtime audio-language model
Input	Speech, text, and images
Output	Speech and text
Context	32k tokens
Max output	4k tokens
Tools	Search knowledge, inspect sources, and navigate related records
Turn taking	Server-side voice activity detection

Configuration

Setting	How it is configured
Where it runs	Observatory call mode. A live session, not a convert job.
Noise	Near-field reduction is on by default.
Serving	AiQL selects the checkpoint and the speaking voice.

⌘I