voice--v0.0.4

This release significantly improves wake-word detection reliability: a pre-roll buffer prevents clipping the start of utterances, a sliding-window fuzzy matcher handles filler words and per-word edit tolerance, Whisper is biased toward your configured trigger vocabulary, and loudness is normalized before transcription to help quiet Bluetooth/HFP captures.

Changed

• dispatcher: trigger matching — sliding window + per-word tolerance — the matcher now tries up to three starting offsets so filler words prepended by Whisper ("um hey jarvis") don't prevent a match. Tolerance is now one allowed edit per trigger word (previously a character-count heuristic), making it more predictable across short and long triggers.
• dispatcher: default Whisper model tiny → base — better transcription accuracy out of the box at a modest CPU cost. Override with whisper.model: tiny in config.yaml if needed on low-power hardware.

Added

• dispatcher: pre-roll audio buffer — 320 ms of audio is buffered before VAD onset so the first syllable of an utterance is never clipped before Whisper sees it.
• dispatcher: Whisper initial-prompt biasing — the configured trigger phrases are fed to Whisper as an initial_prompt, nudging the model toward your vocabulary and reducing mis-transcriptions of trigger words.
• dispatcher: normalize_gain config option — loudness-normalizes captured audio before transcription (default true). Particularly effective for AirPods and other Bluetooth sources that capture at low gain. Disable with audio.normalize_gain: false in config.yaml.
• dispatcher: trigger_tolerance config option — override the per-word edit tolerance globally (audio.trigger_tolerance: N in config.yaml). Leave unset to use the automatic one-edit-per-word default.

Upgrade Instructions

Upgrade dispatcher only — no plugin changes in this release. No re-pairing required. The default Whisper model changes from tiny to base; if transcription feels slower on your hardware, add whisper: {model: tiny} to config.yaml to revert.