Replaces fuzzywuzzy intent matching with ONNX MiniLM-L6-v2 sentence
embeddings (cosine similarity, threshold 0.45). Now triggers on
semantically equivalent phrases not literally in commands.yaml.
Adds optional noisereduce preprocessing before Vosk (off by default,
DENOISE_ENABLED config flag). Inference latency: 2.5ms median CPU.
Single window (900x720, dark background) loading web/index.html with a
Python bridge exposing list_existing_commands, list_exes,
list_sound_names, preview_sound, slugify, validate_draft, save_command,
plus llm_suggest hook for M3. Sound preview spawned on a daemon thread
so the GUI does not block.
ruamel.yaml is needed by tools/command_builder/yaml_writer to round-trip
commands.yaml without losing comments. pyautogui drives the new keys
action. pywebview is added now too so M2's GUI bootstrap can land
without another deps bump.
shell uses subprocess (with shell=True so user can write a normal
command line); url uses webbrowser.open; keys uses pyautogui.hotkey
for combinations or pyautogui.typewrite for raw text. pyautogui is
imported lazily so the bot still starts on machines without a GUI.
Uses ruamel.yaml so existing comments / formatting in commands.yaml
survive a save. Each write creates commands.yaml.bak first, then
atomically renames a .tmp file into place. Tests cover round-trip for
all action types plus comment-preservation on append.
Adds webrtcvad-based smart command-end detection (fixed 10s window
replaced by VAD with configurable silence/min/max thresholds) and
migrates execute_cmd to a YAML-driven plugin system with typed
actions (exe, play_sound, system, multi, sleep). 17 commands
migrated; all original behavior preserved.
All 17 voice commands migrated from a flat phrase list to the
phrases/action shape. Behaviour preserved 1:1 with the previous
elif chain — including the 200ms post-Popen pause for music_*
controls and the 500ms pause before the 'ready' chime on the
audio-device switch commands.
Replaces the 100-line elif chain with a small dispatcher. Each entry in
commands.yaml now carries both 'phrases' (existing trigger list) and
'action' describing how to run it. Action types: exe, play_sound,
system (volume_mute/unmute/exit), sleep, multi.
The data migration lives in the next commit; the dispatcher in this
commit assumes the new schema.
Replace the fixed 10s post-wake-word listen window with a VAD-driven
loop. Once at least 500ms of speech have been seen, 1.2s of silence
closes the window. Hard cap at 15s as a safety net.
Tunable in config.py via VAD_AGGRESSIVENESS, COMMAND_END_SILENCE_MS,
COMMAND_MIN_SPEECH_MS, COMMAND_MIN_LISTEN_MS, COMMAND_MAX_LISTEN_MS.
- Names the assistant as J.A.R.V.I.S. (acronym + MCU pre-Age-of-Ultron)
- British butler tone, addresses user as «сэр»
- Brief replies (1-3 sentences)
- No prudish censorship
wake-word is now handled by vosk, so the picovoice access key is no longer
needed. config.py already had PICOVOICE_TOKEN removed in the previous commit.
uses a second KaldiRecognizer with a JSON grammar of WAKE_WORDS + [unk] for
the wake phase, then hands off to the full-vocab recognizer for the 10s
command window. drops the pvporcupine dependency entirely.
Real secrets now live in untracked dev.env (gitignored). dev.env.example
holds placeholders and is the canonical template — copy it to dev.env on
fresh checkout and fill in real keys.
