picrophone

Talk to the pi coding agent with your voice, and have its replies read back to you — fully on-device.

• Dictation (STT): speak and your words become prompts.
• Read-aloud (TTS): the agent's replies are spoken back.
• Pluggable engines: STT and TTS sit behind a common protocol; swap with one command.
• Interrupt anytime: say "stop" (or "shut up") to cut off a reply.
• Mute: say "mute"/"unmute", or run /mute, to pause listening without leaving voice mode.
• Thinking sound: a soft ambient tone plays while the agent works.

It ships as a toggleable pi extension backed by a small native helper.

Quickstart

pi install npm:picrophone

Then run pi, type /voice, and click Allow when macOS asks for Microphone + Speech Recognition (for Picrophone). Now just talk:

• Speak your request — on a short pause it's sent automatically.
• Say "stop" (or "shut up") to cut off a reply.
• Say "mute" / "unmute" (or run /mute) to pause/resume listening.
• Run /voice again to switch to push-to-talk (review before Enter), and once more to turn voice off. A soft chime means it's listening again.

To pin a version: pi install npm:[email protected]. Remove with pi remove npm:picrophone.

Platforms

macOS (supported)

• Apple Silicon or Intel — universal binary (arm64 + x86_64), macOS 13+.
• Dictation: whisper (local WhisperKit, default; models tiny → large) or apple (native SFSpeechRecognizer).
• Read-aloud: av (AVSpeechSynthesizer, default), say, or qwen (on-device Qwen3-TTS).

Configure engines in ~/.pi/agent/picrophone.json (see docs/configuration.md). Full engine and model reference: docs/engines.md.

Windows (coming soon)

Packaging is wired up (an os-gated optional dependency), but no binary ships yet. See docs/platforms.md.

Usage

Toggle with /voice — cycles: off → auto (sends on pause) → push-to-talk (fills the prompt box; press Enter to send) → off. The footer shows the state: 🎙 auto, 🎙 push, 🎙 thinking, 🎙 speaking, 🔇 muted.

Or ask the agent — the extension registers a voice_mode tool, so you can just say/type "turn on voice mode".

Interrupt — while the agent is thinking or speaking, say "stop", "shut up", "cancel", or "wait" to abort the turn and silence the readback. A bare stop word only — "stop the server" is a normal request.

Dictated prompts are prefixed with 🎙 so both you and the model know they came from speech.

Permissions

The first time you start voice mode, macOS prompts for Microphone and Speech Recognition, attributed to Picrophone (not your terminal). Click Allow; review later in System Settings → Privacy & Security. Text-to-speech needs no permission. Why a .app? See docs/architecture.md.

Docs

• Architecture — how the pieces fit together.
• Engines — STT/TTS back-ends and models.
• Configuration — subcommand flags.
• Platform support — macOS/Windows, macOS Tahoe neural TTS, limitations.
• Troubleshooting — logs, common issues, bug reports.
• Contributing — build from source, packaging, releasing.

Feedback

Feature requests and bug reports go in GitHub Issues — see docs/troubleshooting.md for what to include.