Skip to content

Voice messages#

You can speak a message instead of typing it. The microphone in the chat composer records what you say, the runtime transcribes it, and the words go into the thread as your message. The agent answers in the thread, as it does for a message you type.

A voice message produces text and nothing else. The recording is not kept, and the agent speaks nothing back — for a spoken conversation, see voice chat.

Requirements#

A voice message is transcribed on the Realtime endpoint voice uses: managed OpenAI through Lens Agents, an explicit OPENAI_REALTIME_URL and OPENAI_REALTIME_API_KEY, or OPENAI_API_KEY direct to api.openai.com. See Realtime endpoint and credential for the resolution order, and OPENAI_TRANSCRIPTION_MODEL in Voice meetings for the transcriber.

A voice message uses no voice of the agent's own, so it records while a voice chat or a Meet call is live on another connection.

Record a voice message#

To send a voice message:

  1. In the chat composer of the web UI, click the microphone. The browser asks for the microphone the first time.
  2. Speak. The composer shows Recording, the level of your voice, and how long it has been recording.
  3. Click send. The composer shows Transcribing…, and the words are then sent as your message.

Click ✕ while recording to discard the recording. A recording that reaches five minutes is transcribed and sent from there.

One microphone serves the page, so recording and a call take turns: end a live voice chat to record a voice message, and send or discard a recording to place a call. The disabled control's tooltip says which of the two is holding the microphone.

Troubleshooting#

The microphone is blocked. Allow it for the page in the browser's site settings, then click the microphone again.

No speech was heard. The recording held no words. Record again, and speak before pressing send.

The transcript did not arrive. The runtime answered nothing in time. Record the message again.

The connection dropped, so the dictation ended. The recording went with the chat socket. The composer reconnects on its own; record again once it does.