Voice mode

Talk to Syntic Code instead of typing. While voice mode is on, the microphone listens all the time: you speak, end with a send word such as “go go” (or just stop talking for a while), and your words are sent as a message. Permission prompts and questions can be answered out loud, and the start of each reply is read aloud.

Voice mode is off until you turn it on. Speech is turned into text on your computer; no audio leaves it. Only the text is sent, as if you had typed it.

macOS is supported, Linux works on a best-effort basis, and Windows is not supported yet.

Turn it on

Run /voice on. The first time, Syntic Code explains what voice mode does and asks for your agreement; run /voice accept to agree. It then sets up a local speech engine in ~/.syntic/voice:

  • Python 3.10 to 3.13 must be installed;
  • a private Python environment is created there with the speech packages;
  • the first start downloads the speech model once (about 500 MB);
  • macOS asks once for microphone access for your terminal app.

You can also start with syntic --voice, or with the microphone closed: syntic --voice-off. Set "voice": {"enabled": true} in your user settings to start every session with voice on.

⚠️

Always-on means anything said in the room (a TV, a call, other people) can be heard and sent. Say “voice off”, press F8, or run /voice off when you need quiet.

Sending

End what you say with a send word: “Check the login tests again, go go.” The send word is removed from the message. The defaults are go go, finish, do it and send it.

submitModeA message is sent when
both (default)you say a send word, or after silenceMs of quiet (10 seconds)
wordonly a send word; pauses add to a draft
silenceonly after silenceMs of quiet

Nothing is sent while you are speaking, while a reply is being read aloud, while a prompt or question is open, or for two seconds after you press a key. While a reply is running, what you say is queued as your next message.

Everything by voice

ToSay
Send”…go go” / “…finish” / “…do it” / “…send it”
Answer a permission prompt”yes” · “always” · “no” · “no, use pnpm instead” · “option two”
Answer a questionthe option’s name · “option two” · “the second one” · “skip” · “submit”
Stop the running reply”stop” / “interrupt”
Drop the draft and the queue”scratch that”
Pause listening”voice off” (then only “voice on” or “wake up” is heard)
Spoken replies off / on”stop talking” / “talk to me”
Change the send word”set send word to banana” / “add send word banana”

Prompts and questions only take these answers; your words are never typed into a dialog.

Keys and commands

Key / commandWhat it does
F8Open or close the microphone.
F9Spoken replies on or off.
/voice status · /voice settingsWhat voice is doing, and its settings.
/voice replies on|offSpoken replies (saved).
/voice send-words "go go" finishYour send words (saved).
/voice submit word|silence|both · /voice silence 10When messages are sent (saved).
/voice vocab Syntic "Camila Live"Names to spell right (saved).
/voice off · /voice setupTurn voice off; install or repair the engine.

The footer shows listening, recording, transcribing, asleep or voice off, with the draft’s word count, queued messages and whether replies are spoken.

Spoken replies

When a reply finishes, its first two sentences are read aloud (without code, tables, links or paths); the full reply stays on screen. The microphone is muted while it speaks. Syntic Code uses say on macOS and spd-say or espeak on Linux; set SYNTIC_VOICE_TTS to use another program.

Settings

In the voice block of your user settings.json (defaults shown):

{
  "voice": {
    "enabled": false,
    "submitMode": "both",
    "sendPhrases": ["go go", "finish", "do it", "send it"],
    "silenceMs": 10000,
    "vocabulary": [],
    "model": "small.en",
    "language": "en",
    "vadAggressiveness": 2,
    "whileBusy": "queue",
    "speakReplies": true,
    "sounds": true,
    "toggleKey": "f8",
    "speakToggleKey": "f9",
    "inputDevice": null
  }
}

vocabulary lists names the speech model would otherwise mishear. model ranges from tiny.en (fastest) to medium.en (most accurate). vadAggressiveness (0–3) ignores more background noise when higher.

The speech engine

The speech work runs in a separate local program, so it can be replaced: today a small Python program with faster-whisper, installed in ~/.syntic/voice (log: engine.log). Syntic Code and the engine exchange one JSON object per line (config, listen, mute, cut, quit one way; hello, state, partial, final, level, error the other). Set SYNTIC_VOICE_ENGINE to use another engine that speaks the same protocol.

To remove voice completely, delete ~/.syntic/voice and the speech model cache (~/.cache/huggingface).