{
  "name": "wake-word-detector",
  "description": "DNN classifier on Google Speech Embeddings for wake word detection on RPi4",
  "version": "v10",
  "sample_rate": 16000,
  "chunk_samples": 1280,
  "n_mel_bins": 32,
  "mel_window": 76,
  "mel_step": 8,
  "embedding_dim": 96,
  "n_feature_frames": 16,
  "classes": [
    "wake_word",
    "background"
  ],
  "test_accuracy": 0.9978947639465332,
  "quantized_accuracy": 0.998,
  "model_size_bytes": 112944,
  "input_type": "int8",
  "output_type": "float32",
  "input_shape": [
    1,
    16,
    96
  ],
  "output_shape": [
    1,
    2
  ],
  "input_scale": 0.6137130260467529,
  "input_zero_point": -11,
  "architecture": "DNN (Flatten, Dense64, LayerNorm, Dense32, LayerNorm, Dense2)",
  "framework": "TensorFlow/Keras -> TFLite INT8",
  "frozen_models": {
    "melspectrogram": "https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/melspectrogram.tflite",
    "embedding": "https://github.com/dscripka/openWakeWord/releases/download/v0.5.1/embedding_model.tflite"
  },
  "melspec_transform": "x/10 + 2",
  "pipeline": "audio -> melspectrogram.tflite -> embedding.tflite -> DNN classifier",
  "loss": "SparseCategoricalCrossentropy with class_weight upweighting wake_word"
}