feat: real inference pipeline — STT, tone classifier, diarization, mic capture
- cf_voice/stt.py: WhisperSTT async wrapper (faster-whisper, thread-pool executor,
rolling 50-word session prompt for cross-chunk context continuity)
- cf_voice/classify.py: ToneClassifier — wav2vec2 SER + librosa prosody flags
(energy, ZCR speech rate, YIN pitch contour) mapped to AFFECT_LABELS
- cf_voice/diarize.py: Diarizer async wrapper around pyannote/speaker-diarization-3.1;
speaker_at() helper for Navigation v0.2.x wiring
- cf_voice/capture.py: MicVoiceIO — sounddevice 16kHz mono capture, 2s window
accumulation, parallel STT+classify tasks, shift_magnitude from confidence delta
- cf_voice/io.py: make_io() now returns MicVoiceIO when CF_VOICE_MOCK is unset
- cf_voice/context.py: classify_chunk() split into mock/real paths; real path
decodes base64 PCM and runs ToneClassifier synchronously (cf-orch endpoint)
- pyproject.toml: inference extras expanded (faster-whisper, sounddevice,
librosa, python-dotenv)
- .env.example: HF_TOKEN, CF_VOICE_WHISPER_MODEL, CF_VOICE_DEVICE, CF_VOICE_MOCK,
CF_VOICE_CONFIDENCE_THRESHOLD
Prior art ported from: Plex-Scripts/transcription/diarization.py (pyannote
setup), devl/ogma/backend/speech/transcription_engine.py (faster-whisper
preprocessing and session prompt pattern).
fed6388b99