voice-skill

Ask your agent something, say "voice this" (or turn voice mode on), and it writes its answer on screen and dictates almost immediately a listening-optimized version using Kyutai Pocket TTS, a fully local and ultra-lightweight TTS model:

• Runs on CPU with zero config. No remote server, no GPU, no model wrangling. pocket-tts generate --text "hi" works out of the box; the first run pulls the model and you're done.
• Ultra fast time-to-first-word. It's a distilled, streaming model: it starts emitting audio almost immediately instead of rendering the whole clip first, so a reply begins speaking right away. With voice mode the model stays warm and replies start in a couple of seconds.
• Clone any voice. Pick a built-in voice by name (michael, alba, eve, ...) or point --voice at any .wav to clone it. This skill defaults to michael, and you can swap voices with one env var.

How it works

SKILL.md instructions plus a speak.sh helper that wraps Pocket TTS:

• One-off - speak.sh "text" runs pocket-tts generate and plays the result through your speakers. The model reloads each call, nothing stays running.
• Voice mode - speak.sh on launches pocket-tts serve (a local FastAPI server) so the model stays warm. Subsequent replies stream through POST /tts and start playing. speak.sh off stops it.

speak.sh auto-routes: if the server is up it uses it, otherwise it falls back to a one-off generate. Defaults: English, voice michael.

Requirements

• curl on your PATH.

• An audio player on your PATH. The helper auto-detects a common one (afplay, paplay, aplay, ffplay, play, mpv, or cvlc). Override with POCKET_TTS_PLAYER to force a specific one.

• Pocket TTS on your PATH as pocket-tts:

  uv tool install pocket-tts

Install

With the skills CLI (recommended)

This repo is compatible with the open agent skills ecosystem:

npx skills add segudev/voice-skill

That installs the skill (with its speak.sh helper) into your agent's skills directory.

Manual

git clone https://github.com/segudev/voice-skill.git
cp -r voice-skill/skills/voice <your-agent-skills-directory>/voice
chmod +x <your-agent-skills-directory>/voice/speak.sh

(Optional) Allow the helper in your agent's permission settings to avoid a prompt every time it runs.

Usage

Talk to your agent normally and trigger it by voice:

• "use voice to respond" / "voice this" / "say that aloud" - voice the next reply (one-off)
• "voice mode on" - keep voicing every reply via the warm server
• "voice off" / "text only" - stop voicing and shut the server down

You can also drive the helper directly (use the path where the skill was installed):

speak.sh "Hello from Pocket TTS."   # speak text
echo "piped text works too" | speak.sh
speak.sh on                          # voice mode on (warm server)
speak.sh status                      # is the server up?
speak.sh off                         # voice mode off

Configuration

speak.sh reads these environment variables (defaults shown):

Variable	Default	Description
POCKET_TTS_VOICE	michael	A built-in voice name (alba, eve, george, michael, ...) or a path to a .wav/.safetensors clone file.
POCKET_TTS_LANG	English	Language model id, e.g. french_24l, spanish_24l, german_24l, italian_24l, portuguese_24l.
POCKET_TTS_PORT	8000	Port for the warm server.
POCKET_TTS_PLAYER	auto-detect	Audio player command, e.g. POCKET_TTS_PLAYER="ffplay -nodisp -autoexit -loglevel quiet".
POCKET_TTS_STARTUP_TIMEOUT	180	Seconds to wait for the server to become healthy on first start.

For a non-English voice mode, start the server with the matching model:

POCKET_TTS_LANG=french_24l speak.sh on

Notes

• The first run downloads the model (and the selected voice) from Hugging Face and is slower; later runs are quick.
• Voice mode keeps a process running until you call speak.sh off.

Credits

• Kyutai Pocket TTS does the actual text-to-speech.
• This repo is just the skill glue around it, packaged for the skills ecosystem.