Before you begin
- Who: any user of the web chat. No setting to turn on.
- Where: the chat input of Chat (home screen and conversations), the Agent Factory playground, the embedded chat widget when its owner enables dictation, and the Ask bar of a Collection.
- Browser: any modern browser that can record audio: Chrome, Edge, Brave, Firefox, or Safari. Dictation also works behind corporate proxies.
- Requires: a microphone, and an organization where a transcription model is available to you. If the microphone button is missing, see The microphone button is missing.
Dictate a message
1
Start recording
In the chat input, select the microphone button (Dictate). The first time, your browser asks for permission to use the microphone: allow it.While the browser prompt is open, the button shows Waiting for microphone access…. Select it again to cancel the request.
2
Speak
Say your message. You can record for up to 2 minutes: at 2 minutes, recording stops on its own and what you said is transcribed.
3
Stop recording
Select the microphone button again. The button shows a spinner and Transcribing… while your audio is turned into text.
4
Review and send
The transcribed text is added to the input, after any text you had already typed. Correct it if needed, then send it as you would any typed message.
Verify the result
- Type a few words in the chat input, then dictate a short sentence.
- Confirm that the sentence appears after your typed words and that no message was sent.
- Send the message and confirm that the agent answers it like any typed message.
How dictation works
Your browser records the audio and sends it to the Prisme.ai platform, where your organization’s transcription model turns it into text. The browser’s own speech service is not used, which is why dictation works the same way in every modern browser and behind corporate proxies.- Your audio is never stored. It is processed in memory and discarded once transcribed. It is not saved in storage, in the activity feed, or in logs. The transcribed text is then handled exactly like text you type.
- Language. The chat sends your interface language to the transcription model as a hint, so that short dictations such as “Yes” or “Oui” are recognized correctly. In the embedded chat widget, the model detects the language on its own.
- Mobile apps. The Prisme.ai mobile apps are not affected: they keep using your device’s speech recognition.
For administrators
Dictation relies on a transcription model in the LLM Gateway model catalog. The microphone button only appears for a user when all of these are true:- the platform or the organization has a transcription model, and the organization’s allowed models include it when the organization restricts models;
- the user’s role grants the
llm-gateway:models:audio-transcriptionspermission (Use Audio Transcriptions), which the system Member, Agent Maker and Builder roles include by default. A custom role must have it added.
POST /v1/audio/transcriptions reference to call the same transcription service from your own code.
Troubleshooting
The microphone button is missing
Cause: one of these applies:- your browser cannot record audio, for example an old browser or a page embedded without microphone access;
- your organization has no transcription model available, or its allowed models do not include one;
- your role does not grant the
llm-gateway:models:audio-transcriptionspermission; - in an embedded chat widget, you are not signed in: anonymous sessions do not hold this permission.