Voice input¶

Speech-to-text for the reyn chat inline CUI, powered by faster-whisper. Revived (#4187) as a reimplementation against the current CUI — the retired Textual TUI this was originally built for (src/reyn/chat/tui/) was deleted wholesale when reyn chat moved to the inline CUI, so the STT core (record → transcribe) is ported near-verbatim from that history but the key binding and composer wiring are new.

What it does¶

Press F2 to start recording; press F2 again to stop and transcribe. The transcribed text is inserted at the composer's cursor, ahead of any existing draft — it is never auto-submitted, so the operator reviews (and can edit) before pressing Enter. Esc while recording cancels it without transcribing. A recording left open past voice.max_duration_s (default 300 s) auto-stops and transcribes on its own, the same as a manual F2 press, so a forgotten open mic does not run forever.

F2, not the retired TUI's Ctrl+R: #4187 measured Ctrl+R as colliding with reverse-history-search, a terminal-wide convention users bring to any text-input surface (the Composer is a multi-line text area) — F2, the retired TUI's own alias, carries no such conflict.

Backend¶

faster-whisper (Systran/faster-whisper-<model> on HuggingFace) provides local, offline transcription via CTranslate2. The model downloads on first use and is cached in the HuggingFace hub cache — the FIRST F2-stop after launch pays that cost, not the press that starts recording. Audio is captured at 16 kHz mono via sounddevice (PortAudio). Inference runs in asyncio.to_thread so the Textual event loop never blocks.

Enabling¶

pip install "reyn[voice]"

The base install never imports sounddevice or faster-whisper. Without the extras, pressing F2 shows a friendly in-conversation message naming the extra to install, rather than doing nothing or crashing.

Configuration (`reyn.yaml`)¶

voice:
  enabled: true           # false = hard-disable F2 even if deps are installed
  model: small             # tiny | base | small | medium | large-v3
  language: "ja"           # ISO 639-1 code; "" or null = Whisper auto-detect
  device: cpu               # cpu | cuda (no Metal backend; explicit cpu avoids Mac issues)
  compute_type: int8        # int8 | float16 | float32
  cpu_threads: 4             # pin to 4 to avoid OpenMP deadlock on Apple Silicon
  num_workers: 1              # parallel transcribe streams — 1 keeps memory/threads low
  sample_rate: 16000            # Whisper expects 16 kHz mono — leave this alone
  max_duration_s: 300.0        # auto-stop+transcribe a forgotten recording after this long

Language detection¶

Default is language: "ja" (Reyn's Japanese-enterprise focus). Short clips misidentify as other languages at non-trivial rates when auto-detect is on. Set language: "" or language: null to opt back into Whisper auto-detection.

Limitations¶

No Metal backend. faster-whisper does not support Apple Metal/MPS; device: cpu is correct on Mac.
Model download on first use. small is ~460 MB.
Silence gate. Audio with peak amplitude below 0.005 is skipped entirely (never reaches the model) — Whisper hallucinates on pure noise. Textual's own VAD filter is disabled for the same reason (it rejects quiet-but-real speech more often than it rejects noise).
No dictate-and-send, no debug-WAV dump. The retired TUI's later additions (Enter-while-recording stops+submits in one keystroke; a REYN_VOICE_DEBUG env var that saves captured audio to disk) were not ported — #4187 ported the STT core and rebuilt only the interface glue the current CUI needs. Revisit if wanted; nothing here blocks adding them.