> For the complete documentation index, see [llms.txt](https://oodlz-ai-studio.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://oodlz-ai-studio.gitbook.io/docs/inbound-agent/every-channel/voice/speech-to-text.md).

# Speech-To-Text

Audio binary in, transcript out — Whisper transcription wired to the main agent pipeline with shared memory.

> **Speak the question; the agent hears it.**

Voice input is handled via a dedicated transcription pipeline. Audio binary is received from the voice client, validated, transcribed by Whisper, and the resulting transcript flows into the main agent pipeline with the same session reference as text turns.

## How It Works

The voice client captures audio in the browser or app, packages it as a binary payload, and POSTs it to the transcription endpoint. The pipeline validates the payload — confirming a usable audio binary is present and standardising the field name expected by Whisper — and calls Whisper for the transcription. The transcribed text is returned to the client and continues into the main agent pipeline carrying the visitor's session reference, so the spoken question becomes part of the same conversation memory as any text turn before or after it.

If transcription fails or returns empty — silent recording, unrecognisable audio, network drop — a user-friendly fallback is returned asking the visitor to try again. The agent is not invoked on failed transcripts; cost is incurred only when there is something for the agent to actually answer.

## What You Get

* Whisper transcription from any voice client connected to the inbound agent.
* One conversation thread spanning voice and text turns of the same session.
* Graceful empty-audio handling — no agent calls fired on noise.