Voice Agent Legacy
Voice input in this panel does not work
The server endpoint the panel needs, /ws/voice/, is commented out in the backend as unmaintained, and the frontend no longer opens that socket on load. With no connection, the microphone and wake-word buttons stay permanently disabled, and the panel’s speech recognition never starts.
For dictation today, use the microphone button in the Composer.
The Voice Agent panel was built to take spoken commands, transcribe them, and send them to the legacy chat backend. The transcription half of it is inert. The typing half still works, and behaves like a smaller version of the legacy chat panel.
Where to find it
View > Agents > Voice Agent. It is not included in any layout preset.
What still works
- The text box. A free-text area at the top of the panel. Type into it and press Send.
- Enable Visual Brainstorming / Disable Visual Brainstorming. Toggles a flag sent with the message.
- The mode dropdown. Narrator, Agent, or Coder. Coder marks the message as a code request. Narrator and Agent send an identical payload; the only difference between them is which socket the Send button checks before enabling itself.
- Send. Submits the text.
Sending clears the context points on the map first, then looks at what your text refers to:
- If you have a single point selected in point mode, the message is sent with that point attached.
- Otherwise, every bag or cluster whose name appears in your text is sent as its own message, carrying that group’s point ids.
- Otherwise, your selected context items are used the same way.
- Otherwise, the message is sent with no attached points.
What does not work
- The microphone button. Disabled, because the voice socket never connects.
- The wake-word On/Off toggle. Disabled for the same reason.
- Speech recognition. The panel only starts listening once the server tells it to, which cannot happen.
What voice input actually is in Mantis
Where speech input does work, in the Composer’s mic button, it is the browser’s own Web Speech API. That means:
- English only. The recognition locale is hardcoded to
en-US, with no selector. - Dictation only. Speech becomes text in the input box. Nothing is interpreted as a command by the speech layer itself.
- It never speaks back. There is no speech synthesis in the voice path, no voices to choose from, and no spoken responses.
- No speaker identification or voice profiles. Nothing about who is speaking is captured.
- Browser dependent. In a browser without the Web Speech API the mic button renders disabled, and where it works, the recognition quality is the browser’s, not Mantis’s.