Voice: finalise on faster-whisper + energy wake, remove Vosk
Full-day voice-stack refactor. Experiments run and reverted:
- Gemini Live HTTP microservice (Python 3.8 env incompat, latency)
- Vosk grammar STT (English lexicon can't decode 'Sanad'; big model
cold-load too slow on Jetson CPU)
Kept architecture:
- Voice/wake_detector.py — pure-numpy energy state machine with
adaptive baseline, burst-audio capture for post-hoc verify.
- Voice/marcus_voice.py — orchestrator with 3 modes
(wake_and_command / always_on / always_on_gated), hysteretic VAD,
pre-silence trim (300 ms pre-roll), DSP pipeline (DC remove,
80 Hz HPF, 0.97 pre-emphasis, peak-normalize), faster-whisper
base.en int8 with beam=8 + temperature fallback [0,0.2,0.4],
fuzzy-match canonicalisation, GARBAGE_PATTERNS + length filter,
/s-/ phonetic wake-verify, full-turn debug WAV recording.
Config-driven vocab (zero hardcoded strings in Python):
- stt.wake_words (33 variants of 'Sanad')
- stt.command_vocab (68 canonical phrases)
- stt.garbage_patterns (17 Whisper noise outputs)
- stt.min_transcription_length, stt.command_vocab_cutoff
Command parser widened (Brain/command_parser.py):
- _RE_SIMPLE_DIR — bare direction + verb+direction combos
('left', 'go back', 'move forward', 'step right', ...)
- _RE_STOP_SIMPLE — bare stop/halt/wait/pause/freeze/hold
- All motion constants sourced from config_Navigation.json
(move_map + step_duration_sec) via API/zmq_api.py; no more
hardcoded 0.3 / 2.0 magic numbers.
API/audio_api.py — _play_pcm now uses AudioClient.PlayStream with
automatic resampling to 16 kHz (matches Sanad's proven pattern).
Removed:
- Voice/vosk_stt.py (and all Vosk references in marcus_voice.py)
- Models/vosk-model-small-en-us-0.15/ (40 MB model + zip)
- All Vosk keys from Config/config_Voice.json
Documentation synced across README, Doc/architecture.md,
Doc/pipeline.md, Doc/functions.md, Doc/controlling.md,
Doc/MARCUS_API.md, Doc/environment.md changelog.
Known limitation: faster-whisper base.en on Jetson CPU + G1
far-field mic yields ~50% command-transcription accuracy due
to model capacity and mic reverberation. Wake + ack + recording
+ trim + Whisper + fuzzy + brain + motion all verified working
end-to-end. Future improvement path (unused): close-talking USB
mic via pactl_parec, or Gemini Live via HTTP microservice.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
kassam
5d839d4f4e