ovos-ww-plugin-efficientword-net

The model, the reference files and the inference method are from
EfficientWord-Net (https://github.com/Ant-Brain/EfficientWord-Net),
Copyright 2021 ANT-BRaiN, Apache License 2.0.

Changes from EfficientWord-Net:
- models/resnet_50_arc_qint8.onnx is EfficientWord-Net's quantized model
  "slim_93%_accuracy_72.7390%_qint8.onnx", renamed.
- references/*_ref.json are EfficientWord-Net's sample references, unchanged.
- inference.py computes the same 64-band log mel features with numpy only.
  The filterbank code follows python_speech_features
  (https://github.com/jameslyons/python_speech_features), MIT License.
- Reference recordings are centered in the 1.5 s window instead of placed
  at a random offset, so the same recordings always give the same
  references.
- The relaxation time after a detection is counted in audio, not in wall
  clock time.
