Voice input and dictation

Voice is two independent halves. Talking to it is speech recognition (our ASRx module), done by your browser or by a transcription service. It talking back is speech synthesis (Speex), done by one of five voice engines. They are configured separately and fail separately, so one can work while the other does not.

Dictating into the composer

The microphone button runs one of two engines, and its icon shows which one a click starts:

Long take is picked for you as soon as a Deepgram or OpenAI transcription key exists, Quick otherwise. Hover the microphone on desktop, or tap it on mobile, to switch; the choice sticks until you change it. On mobile, picking an engine also starts the recording.

Ctrl + M starts and stops the microphone. While a take is live the shortcut bar swaps in these:

Quick stops itself after a silence, its length set by Mic Timeout at Settings › Voice › Input (5 seconds by default). Long take has no silence detection, so you stop it yourself. You can keep typing while it transcribes: the transcript lands at the cursor, and the recording is then deleted. If the service refuses, one retry runs. If that fails too the audio is kept: Transcription failed - the recording is saved in the Ramble library, transcribe it from there.

The same microphone sits on the Scratchpad and on the Beam inputs.

On Open the microphone is the browser's own recognition, with live text as you speak. A Voice Continuation button re-arms it after each send. There is no long take there - no switcher, no Scratchpad or Beam microphone - and the live-take keys are Ctrl + M for Mic · Send and Esc for Mic · Stop. Nothing on that build uses the transcription settings yet.

Ramble

Ramble records one take of up to 60 minutes and transcribes the whole file free - where to start one, what a reload costs, and what is kept: an hour of speech, transcribed whole.

Holding a spoken conversation

A call is recognition, model generation, and synthesis - three phases in turn. Latency is higher than a realtime voice model; in exchange the call runs on the chat models and the data you already have. Switch to the Calls app, or press the call button in a chat or on an open persona. The call carries that chat's context, or starts with none.

The calls app reports on three things before it dials:

The chat is empty. Calls are effective when the caller has context.
Speech recognition should be good to go.
Voice synthesis should be ready.

all three green, in the footer:   Ready, Set, Call

In the call: Push To Talk while you speak, or leave it listening. Two spoken commands are built in - Stop. makes the assistant hold its answer, Goodbye. ends the call. Ctrl + M still toggles the microphone.

Starting to talk stops the assistant talking. The microphone and the speaker are arbitrated over one audio resource, and the microphone wins. Hanging up cuts playback the same way.

Choosing a voice - synthesis

Five engines synthesize speech, configured in Settings: some run on the AI service keys you already added, some take a key of their own. One engine is active at a time, or Auto picks.

Each engine ships with a default voice, and the preview button reads the same line every time - Hello, this is my voice. A voice is stored as an engine plus its settings.

Pin a voice to a persona and every chat with that persona speaks in it. That includes the line it opens a call with: what a persona is. Outside calls, replies stay silent until Speak Chats at Settings › Voice › Output says otherwise:

Speak Cost Guard caps what reaches the voice engine at 4096 characters, about three minutes of speech. It is on by default. One reply can always be spoken from its message menu's Speak.

All five engines exist in every version (optional services).

Where the audio goes, and what is kept

Whether recognition runs on the device or at the browser vendor's service is the browser's choice, not Big-AGI's. The full account of what leaves your machine, and by which route: what leaves your browser.

Browsers and devices

The rows below are about Quick, the browser's own recognition, and about calls, which use the same engine.

Brave is blocked deliberately, because it exposes the interface and then returns nothing. Long take records with the ordinary media APIs, so it runs in Brave and in every browser above. Synthesis has no restriction either: any browser can play the audio back.

Troubleshoot - audio errors

Related

• An hour of speech, transcribed whole
• What a persona is
• What leaves your browser
• Attachments and outputs
• Optional services