Voice

Voice replies (Pocket TTS)

Dictate your reply out loud with the local Kyutai Pocket TTS install. Defaults: English, voice michael.

How to respond

1. Answer the user's actual question normally in text. The on-screen answer stays.
2. Build a spoken version of that answer optimized for listening, then play it.

The spoken version is NOT the markdown answer read verbatim. Convert it to clean prose:

• Plain sentences only. No markdown, bullets, headings, asterisks, or backticks.
• Drop code blocks, long file paths, and raw URLs - describe them instead ("the config file", "the GitHub link"). Read short identifiers naturally.
• ASCII only. Expand symbols the way a person would say them.
• Keep it tight. Speak the substance, not throat-clearing. If the answer is long, speak a spoken summary and tell the user the full detail is on screen.

Playing it

The speak.sh helper is bundled in this skill's directory. Invoke it by its path (wherever the skill was installed).

Pass the spoken text to the helper:

speak.sh "your spoken text here"

For multi-line or long text, pipe it instead to avoid quoting issues:

speak.sh <<'EOF'
your spoken text here
EOF

The helper auto-routes: if voice mode is on it uses the warm background server (fast); otherwise it does a one-off pocket-tts generate. Either way it plays the audio through your speakers and blocks until playback finishes. Run it after you have written the text answer.

One-off vs voice mode

• One-off (user says "voice this" once): just call speak.sh "...". No server, slower first synth, nothing left running.
• Voice mode (user says "voice mode on" / "keep responding with voice"): first run speak.sh on to launch the warm server, then voice every subsequent reply with speak.sh "..." (now fast). The first on may take a minute to load the model.
• Turn it off (user says "voice off", "stop the voice", "text only"): run speak.sh off and go back to text-only replies.

speak.sh status reports whether the server is up.

Options

The helper reads these environment variables (defaults shown):

• POCKET_TTS_VOICE (default michael) - a built-in voice name (e.g. alba, eve, george, michael) or a path to a .wav/.safetensors clone file.
• POCKET_TTS_LANG (default English) - language model id, e.g. french_24l, spanish_24l, german_24l, italian_24l, portuguese_24l. For French TTS, prefix POCKET_TTS_LANG=french_24l on the command (start the server with it too if using voice mode).
• POCKET_TTS_PORT (default 8000) - server port.
• POCKET_TTS_PLAYER (default auto-detect) - audio player command. The helper auto-detects a common player; set this to force a specific one or if none is found, e.g. POCKET_TTS_PLAYER="ffplay -nodisp -autoexit -loglevel quiet".

Notes

• An audio player must be available. If playback reports "no audio player found", install one (e.g. ffmpeg for ffplay) or set POCKET_TTS_PLAYER.
• The first run downloads the model and selected voice; later runs are quick.