feat: always-on wake word capture with unified audio pipeline

[kocendavid]

Summary

• Adds wake word detection that runs continuously regardless of dictation state — when idle it listens for the wake phrase to start dictation; during dictation it listens for the finish phrase to stop
• Replaces the previous design where wake word capture stopped during dictation and audioManager used timeslice chunks (which produced broken WebM files without headers)
• Uses a stop-start MediaRecorder cycle (3-second chunks) to produce complete, self-contained WebM files that FFmpeg can reliably parse

How it works

State	Behavior
Idle	useWakeWordCapture sends 3-sec audio chunks to main process → wakeWordManager transcribes with base Whisper model → fuzzy-matches against wake phrase → triggers dictation
Dictation starts	Wake word capture continues running (switches to stop-word mode via wakeWordRecordingState IPC)
During dictation	Chunks checked against finish phrase instead of wake phrase
Dictation stops	Seamlessly switches back to wake phrase detection (no restart needed)

Architecture

[Renderer]                          [Main Process]
useWakeWordCapture                  wakeWordManager
  │ 3-sec MediaRecorder chunks        │ separate WhisperServer (base model)
  │ (stop-start cycle = valid WebM)   │ fuzzy phrase matching (Levenshtein)
  └──► wakeWordCheckChunk IPC ──────► └──► transcribe → match → trigger

Key changes

New files

• src/helpers/wakeWordManager.js — Main process module: manages a dedicated WhisperServer instance, handles phrase matching with fuzzy matching, auto-downloads the base model
• src/hooks/useWakeWordCapture.js — Renderer hook: always-on mic capture using stop-start MediaRecorder cycles, polls wake word status, sends chunks via IPC

Modified files

• src/App.jsx — Calls useWakeWordCapture() (no args, runs continuously), notifies main process of recording state
• src/helpers/audioManager.js — Removed finish-phrase timeslice code (~20 lines) — the always-on hook handles stop-word detection now
• main.js — Initializes WakeWordManager, registers IPC handlers, auto-starts if enabled
• preload.js — Exposes wake word IPC API (toggle, set phrases, check chunks, status)
• src/helpers/ipcHandlers.js — Routes wake word IPC calls to manager
• src/helpers/environment.js — Persists WAKE_WORD_ENABLED, WAKE_WORD_PHRASE, WAKE_WORD_FINISH_PHRASE
• src/helpers/windowConfig.js — Disables backgroundThrottling so capture works when window is hidden
• src/hooks/useAudioRecording.js — Strips finish phrase from transcribed text
• src/components/SettingsPage.tsx — Wake word settings UI (enable/disable, set phrases, live detection log)
• src/components/SettingsModal.tsx — Adds Wake Word nav item

What this fixes

The previous approach had a WebM header bug: during dictation, audioManager used MediaRecorder.start(3000) timeslice to send chunks for stop-word detection. Timeslice chunks after the first lack WebM headers, so FFmpeg couldn't parse them. The new design avoids this entirely — useWakeWordCapture uses a stop-start cycle that always produces complete WebM files.

Test plan

• Enable wake word in settings, set wake phrase to "whisper"
• Say "whisper" → recording starts, wake word capture keeps running (no gap in Listener Output)
• Set finish phrase to "done" → say "done" during dictation → dictation stops, text pastes
• During dictation, Listener Output shows [FINISH] mode (not [WAKE])
• Hotkey still works independently for start/stop
• Main transcription produces correct text (single complete blob, no timeslice artifacts)
• Disable wake word → capture stops, no microphone indicator

[kocendavid]

Hey I think this is cool. Try and look if you like it, I will add also differnet words to finish to send or to cancel translation. Maybe some of the stuff is included in the inteligence tab