Wake Word — testing - Cyberwave Docs

Detection threshold

Production and offline tests use the same rule: score ≥ workflow threshold (default 0.5). The Wake Word node exposes one Threshold field in the editor — it applies to all selected models (alexa, hey_jarvis, etc.), not per model. Pass --threshold to offline tools to mirror the value saved on the node:

# Match a workflow node set to 0.35 for hey_jarvis
python manage.py test_wake_word_wav clip.wav --model hey_jarvis --threshold 0.35

Model (fixture)	Typical peak on committed WAV	Often works at default 0.5?
`alexa`	~1.0	Yes
`hey_mycroft`	~0.68	Yes
`hey_jarvis`	~0.41	Often no — try 0.35–0.40

Offline WAV test (edge worker)

Use a recording that contains the wake phrase (e.g. someone saying “alexa”). Copy into the worker container and run the bundled script:

docker cp scripts/test_wake_word_wav.py cyberwave-worker-<ENV_UUID>:/tmp/
docker cp alexa-sample.wav cyberwave-worker-<ENV_UUID>:/tmp/

docker exec -it cyberwave-worker-<ENV_UUID> python3 /tmp/test_wake_word_wav.py \
  /tmp/alexa-sample.wav --wake-words alexa --threshold 0.5

Success (management command on compile host):

  HIT @ 1.28s scores={'alexa': 0.512} (threshold=0.5)
Done: frames=... max_score=0.9998 @ ...s hits_above_threshold=1

On a live edge worker, look for:

Wake word detected: 'alexa' scores={'alexa': 0.512} threshold=0.50, switching to AWAKE

The script:

Resamples any WAV rate/channels to 16 kHz mono int16
Scores 80 ms frames (1,280 samples)
Uses the same Model(wakeword_models=[...]) keys as production

Django management command (compile host)

If openwakeword is installed on the backend container:

docker compose -f local.yml exec django python manage.py test_wake_word_wav \
  src/app/tests/fixtures/audio/alexa_sample.wav --model alexa --threshold 0.5

docker compose -f local.yml exec django python manage.py test_wake_word_wav \
  src/app/tests/fixtures/audio/hey_mycroft_sample.wav --model hey_mycroft --threshold 0.5

docker compose -f local.yml exec django python manage.py test_wake_word_wav \
  src/app/tests/fixtures/audio/hey_jarvis_sample.wav --model hey_jarvis --threshold 0.5

Committed fixtures under cyberwave-backend/src/app/tests/fixtures/audio/ cover alexa, hey_mycroft, and hey_jarvis.

Pytest (test_wake_word_wav_detection.py, @pytest.mark.integration): asserts each fixture reaches an expected peak score (not that 0.5 always fires for every model).
If hey_jarvis at --threshold 0.5 prints No frame reached threshold, lower --threshold to match your workflow node or inspect max_score in the summary line.

Live workflow checklist

Check	Expected
Audio Track buffer	Wake Word (80 ms) — codegen shows `_CwAudioFifoBuffer(1280)`
Worker init log	`wake_words=['alexa'] prediction_keys=['alexa']`
Say wake phrase	`Wake word detected: 'alexa' scores={'alexa': 0.5+}`
Send Alert downstream	Alert name = detected wake word
Assistant signaling	`start_assistant` / `stop_assistant` on twin MQTT topic

Common failures

Symptom	Cause	Fix
No detection, scores always 0	Stale worker using `alexa_v0.1` prediction keys	Recompile + sync
Scores rise but never cross threshold	Threshold too high for that model (common with `hey_jarvis` at 0.5)	Lower Detection Threshold on the node; confirm with `test_wake_word_wav` and `max_score`
No logs from wake word node	Audio Track 4 s buffer	Set Wake Word (80 ms) buffer
`Could not find pretrained model hey_siri`	Removed model still in saved params	Re-save node; pick a built-in word
Compile error: openwakeword not installed	Django image lacks ml-wakeword	Install on compile host or compile from CI with deps
No assistant sounds	Twin not set on Wake Word node or FE not on environment	Enable signaling + select microphone twin

Log levels

# docker-compose or worker env
CYBERWAVE_WORKER_LOG_LEVEL=INFO   # detection + periodic score heartbeat
CYBERWAVE_WORKER_LOG_LEVEL=DEBUG  # per-frame scores when best ≥ 0.15

Regression tests (repo)

Backend (Docker):

cd cyberwave-backend
docker compose -f local.yml exec -T django pytest \
  src/app/tests/test_wake_word_engine_emitter.py \
  src/app/tests/test_wake_word_engine_inputs_outputs.py

Frontend:

cd cyberwave-frontend
pnpm test:run __tests__/openwakeword-models.test.ts __tests__/workflow-audio-chain.test.ts

Integration WAV test (requires openwakeword in test env): test_wake_word_wav_detection.py (marked @pytest.mark.integration).

Concepts

Platform Features

Cyberwave Edge

Technical Reference

Use-Case Recipes

Wake Word — testing

Detection threshold

Offline WAV test (edge worker)

Django management command (compile host)

Live workflow checklist

Common failures

Log levels

Regression tests (repo)

​Detection threshold

​Offline WAV test (edge worker)

​Django management command (compile host)

​Live workflow checklist

​Common failures

​Log levels

​Regression tests (repo)

Detection threshold

Offline WAV test (edge worker)

Django management command (compile host)

Live workflow checklist

Common failures

Log levels

Regression tests (repo)