ovos-ww-plugin-wakeforge

An OpenVoiceOS wake-word plugin. It runs wake-word models made with wakeforge. The models download from Hugging Face on first use.

Install

pip install --pre ovos-ww-plugin-wakeforge

Try it in your browser

Everything runs locally on your machine.

• WakeHuBERT wake words: every ready model, on your microphone or an uploaded file.
• WakePhoneHuBERT: spot a keyword you type, with no trained model.

Try it on your laptop

The demo listens to your microphone. It runs a ready model next to openWakeWord, microWakeWord and Precise, so you can compare them. Use --help for the options.

pip install --pre "ovos-ww-plugin-wakeforge[demo]"
ovos-wakeforge-demo --wakeword hey_mycroft

Use a ready model

Add this to mycroft.conf:

{
  "listener": {
    "wake_word": "hey_mycroft"
  },
  "hotwords": {
    "hey_mycroft": {
      "module": "ovos-ww-plugin-wakeforge",
      "model": "wakehubert_hey_mycroft",
      "listen": true
    }
  }
}

Each ready model is named wakehubert_<word>. The model card lists all words, thresholds, test results and training data: https://huggingface.co/OpenVoiceOS/wakehubert-wakewords.

Threshold

Calibrated models make threshold a sensitivity setting. A higher value gives fewer false activations and fewer detections. A lower value gives more of both. Each model has its own default threshold. Set "threshold" in the hotword config to change it.

WakePhoneHuBERT

WakePhoneHuBERT is a second featurizer, named wakephonehubert-int8. It adds a voice-activity head and an IPA phoneme head to the WakeHuBERT trunk. Models trained on it are named wakephonehubert_<word>. They name their featurizer in their own metadata, so model is the only key you need. The plugin downloads the featurizer on first use.

Zero-shot wake words

The ovos-ww-plugin-wakeforge-zeroshot engine detects a wake word from its IPA phones alone, with no recordings and no trained model. It scores WakePhoneHuBERT's phoneme output for the phones in ipa, every 80 ms over the last 1.5 s of audio. The word fires when two consecutive windows both reach the threshold (confirm_blocks, default 2), which removes single-window matches inside ordinary speech.

"hotwords": {
  "hey_jarvis": {"module": "ovos-ww-plugin-wakeforge-zeroshot", "ipa": "h eɪ dʒ ɑːɹ v ɪ s"}
}

ipa can be a list of pronunciations. Get the phones once with eSpeak NG, which the default thresholds were calibrated with: espeak-ng -q --ipa -v en-us "hey jarvis". The plugin needs no phonemiser. The default threshold depends on the number of phones and is a starting point: set "threshold" (0 is a perfect match) to tune it. Use a phrase of at least six phones, ideally two words. Short single words are unreliable: train a model for them.

Use your own model

Set model to the classifier ONNX that wakeforge exported. Set featurizer to its feature-extractor ONNX. Both can be local paths or URLs. A model trained on a pretrained featurizer names it in its metadata, so you can leave featurizer out. Train models with wakeforge.

{
  "listener": {
    "wake_word": "hey_jarvis"
  },
  "hotwords": {
    "hey_jarvis": {
      "module": "ovos-ww-plugin-wakeforge",
      "listen": true,
      "featurizer": "~/hey_jarvis/featurizer.onnx",
      "model": "~/hey_jarvis/model.onnx"
    }
  }
}

Config keys

The source documents the other keys (smoothing, patience, window_size, ema_alpha, block_ms, streaming, gru_window, isolated_windows, hidden_dim, vad) in ovos_ww_plugin_wakeforge/__init__.py. URL models are cached under ${XDG_DATA_HOME}/wakeforge/.

Credits

Developed by TigreGótico for OpenVoiceOS.

This project was funded through the NGI0 Commons Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 101135429.

License

Apache-2.0