Enable voice input in `reyn chat`¶

Goal: Dictate into the TUI input bar instead of typing, using local Whisper.

Voice input is shipped as an optional extra. The feature stays off by default so the base install has zero extra dependencies.

1. Install the `voice` extra¶

pip install "reyn[voice]"

This pulls three pure-Python packages:

sounddevice — microphone capture (PortAudio binding)
numpy — audio buffer
faster-whisper — local Whisper inference (CTranslate2 backend)

2. Install the system audio library¶

sounddevice needs PortAudio on the OS layer. pip cannot install this for you.

OS	Command
macOS	`brew install portaudio`
Ubuntu / Debian	`sudo apt install libportaudio2 libsndfile1`
Fedora / RHEL	`sudo dnf install portaudio`
Windows	nothing — bundled in the wheel

Verify:

python -c "import sounddevice; print(sounddevice.query_devices())"

You should see your microphone listed.

3. Grant microphone permission (macOS)¶

The first time you press the record key, macOS prompts your terminal app (iTerm2 / Terminal / Ghostty / Wezterm) to access the microphone. Approve it.

If you missed the prompt: System Settings → Privacy & Security → Microphone → enable your terminal.

4. Configure the model (optional)¶

Edit ~/.reyn/config.yaml (or your project's reyn.yaml / reyn.local.yaml):

voice:
  enabled: true        # set false to hard-disable F2 even if deps installed
  model: small         # tiny | base | small | medium | large-v3
  language: ja         # ISO code, or omit for auto-detect
  device: auto         # auto | cpu | cuda | metal
  compute_type: int8   # int8 (fast) | float16 | float32

Defaults: small / auto-detect language / auto device / int8.

The model is downloaded on first use and cached under ~/.cache/huggingface/. Approximate sizes: tiny 75MB, base 140MB, small 460MB, medium 1.5GB.

5. Use it¶

Inside reyn chat:

Key	Action
`Ctrl+R`	Start recording — `🔴 recording…` appears in the conversation pane
`Ctrl+R` (again)	Stop recording, transcribe, append the result to the input bar (review/edit before sending)
`Enter` while recording	Stop recording, transcribe, send immediately — skips the edit step
`Esc` while recording	Cancel without transcribing
`F2`	Alias for `Ctrl+R` (see note below)

The transcribed text lands in the input bar but is not sent. Review, edit if needed, then Enter to submit.

Why Ctrl+R, not F2?

F2 is intercepted by some terminals and by macOS itself (default "F1, F2 = brightness keys"). If you want F2 to work on macOS, enable System Settings → Keyboard → "Use F1, F2, etc. keys as standard function keys", otherwise hold Fn while pressing F2. Ctrl+R has no such friction and is the recommended binding.

Troubleshooting¶

"PortAudioError: Error querying device"¶

PortAudio not installed at the OS layer. Re-run step 2.

"ModuleNotFoundError: No module named 'sounddevice'"¶

The extra was not installed. Run pip install "reyn[voice]".

First transcription takes 30+ seconds¶

Model is being downloaded. Subsequent transcriptions use the cache.

Transcription is wrong on technical terms¶

Whisper struggles with proper nouns, file paths, and code symbols. Edit the input bar before sending, or pin a larger model:

voice:
  model: medium

CPU usage is high during recording¶

Recording itself is cheap. Inference runs only after you stop. If you have a GPU or Apple Silicon, set device: metal (Apple) or device: cuda (NVIDIA).

I want to disable voice input¶

pip uninstall sounddevice faster-whisper — the voice config block is ignored if the modules are missing.

Privacy note¶

faster-whisper runs fully on-device. No audio is sent to any server. If you later switch to the OpenAI Whisper API for accuracy, that uploads audio to OpenAI — opt-in only, never the default.

Enable voice input in reyn chat¶

1. Install the voice extra¶