voice — TTS & STT tool for AI agents

voice speaks text aloud using Kokoro TTS and transcribes speech using Moonshine STT on Apple Silicon. Use it to talk to your user, listen for their response, or run a full voice conversation loop.

Quick reference

Speak (TTS)

# Speak text (backward compatible — no subcommand needed)
voice Hello, I finished the task.

# Explicit say subcommand with options
voice say -v am_michael "Switching to a male voice."

# Speak from a pipe
echo "Build complete." | voice say

# Read a file aloud (strip markdown first)
voice say --markdown -f README.md

# Save to WAV instead of playing
voice say -o result.wav "Here is your audio."

# Precise pronunciation via IPA phonemes
voice say --phonemes "həlˈO wˈɜɹld"

Converse (speak + listen)

# Speak text, then immediately listen for a response
voice converse "How are you today?"

# With voice and speed options
voice converse -v am_michael -s 1.2 "What do you think about that?"

Listen (STT)

# Record from mic, transcribe on Enter/Ctrl+C
voice listen

# Continuous mode — transcribe segments as you speak, split on silence
voice listen --continuous

# Transcribe a WAV file
voice transcribe recording.wav

JSON-RPC server

# Start the server (for programmatic control)
voice serve -v am_michael

# Speak
→ {"jsonrpc":"2.0","method":"speak","params":{"text":"Hello"},"id":1}
← {"jsonrpc":"2.0","result":{"duration_ms":1800,"chunks":1},"id":1}

# Listen (ding plays, records, auto-stops on silence)
→ {"jsonrpc":"2.0","method":"listen","id":2}
← {"jsonrpc":"2.0","result":{"text":"I heard you","tokens":4,"duration_ms":3200},"id":2}

# Cancel current playback or recording
→ {"jsonrpc":"2.0","method":"cancel","id":3}

# Other methods: set_voice, set_speed, list_voices, ping

When to use

• Get attention: Speak when a long task finishes, a build fails, or you need input
• Read content: Pipe text through voice say to read back docs, errors, or summaries
• Confirm actions: "Deploying to production" before doing something irreversible
• Listen for input: Use voice listen to capture a spoken response from the user
• Voice conversation: Use voice converse to speak then listen in one shot, or voice serve for programmatic control
• Transcribe recordings: Use voice transcribe to convert audio files to text

Tips

TTS tips

• Use -q for quiet mode — suppresses phonemes and progress, only errors print
• For long text, voice automatically chunks at ~510 phonemes and streams playback
• Stderr shows phoneme output — useful for debugging pronunciation
• Use --sub word=replacement to fix names: voice say --sub kubectl=cube-cuddle "Restarting kubectl"
• A .voice-subs file in the project root is auto-discovered for persistent fixes
• Wrap substitution values in /slashes/ for raw phoneme overrides: Kokoro=/kˈOkəɹO/

STT tips

• A ding sound plays when the mic is ready — wait for it before speaking
• Bluetooth mics (AirPods) have ~0.5s latency; the ding helps you time it
• Noise floor is calibrated automatically — works with MacBook mic or AirPods
• Use STT_MODEL=UsefulSensors/moonshine-tiny for faster (but less accurate) transcription
• Default model is moonshine-base (61M params, ~50× real-time on Apple Silicon)

JSON-RPC tips

• voice serve loads the TTS model at startup; STT model loads lazily on first listen
• cancel interrupts the current speak or listen mid-operation
• speak supports per-request voice and speed overrides without changing defaults
• listen params are tunable: noise_multiplier, calibration_ms, silence_timeout_ms
• Notifications (requests without id) are fire-and-forget — no response returned

Subcommands

Builtin voices (no network)

af_heart (default), af_bella, af_sarah, af_sky, am_michael, am_adam, bf_emma

Install

git clone https://github.com/rgbkrk/voice.git
cd voice
cargo install --path crates/voice-cli

Requires macOS with Apple Silicon, Git LFS, and Rust 1.85+. TTS model weights download on first voice say (~312MB, cached). STT model weights download on first voice listen (~246MB, cached).