ovos-ww-plugin-efficientword-net

An OVOS wake-word plugin for EfficientWord-Net, a few-shot hotword detector. You give it 4 or more recordings of a wake word, and it detects that word. It needs no training.

The plugin embeds the last 1.5 s of audio with EfficientWord-Net's ResNet-50 ArcFace model and compares the embedding with the embeddings of your recordings. A detection is a score of 0.7 or more, where the score is (cosine similarity + 1) / 2 against the closest recording.

Install

pip install --pre ovos-ww-plugin-efficientword-net

The plugin needs only onnxruntime and numpy. It carries the quantized EfficientWord-Net model (22 MB) and the upstream sample references, so it works offline and does not install PyAudio.

Configuration

Use one of the bundled wake words: alexa, balloon, computer, lights_off, lights_on, mobile or mycroft.

"hotwords": {
  "computer": {
    "module": "ovos-ww-plugin-efficientword-net",
    "listen": true
  }
},
"listener": {"wake_word": "computer"}

Or give your own recordings: a folder of 16-bit WAV files, one wake word per file, at least 4 files. Any sample rate works.

"hotwords": {
  "hey_computer": {
    "module": "ovos-ww-plugin-efficientword-net",
    "samples": "~/wakewords/hey_computer",
    "listen": true
  }
}

Measured results

Each clip was streamed through the plugin with 1 s of silence before and after it, in chunks of 1024 samples, with the default settings and the bundled references.

300 one-second speech clips are not enough to estimate false activations per hour. Measure on your own audio before you use a wake word.

The quantized model gives scores within 0.005 of the full-precision model on the upstream recordings. One score takes about 50 ms on a laptop CPU, and the plugin scores every hop_seconds. Increase hop_seconds on slow hardware.

Limits

• Detection quality depends on the recordings. Record in the room and with the microphone that the device uses.
• The bundled references are English words.
• A wake word longer than 1.5 s does not fit in the window.

Credits

EfficientWord-Net is by ANT-BRaiN, Apache-2.0. The plugin carries its model, its sample references and a numpy version of its inference code. See NOTICE for the changes.

Report problems at the issue tracker.

Funding

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.