Small Footprint Keyword Spotting with different ML architectures

This repository contains the implementation in Tensorflow 2.11.0 of different models for the KWS task.

CNN with dropout(0.2) - 4 classes (3 keywords + 1) - test acc: 97.72%

• Total params: 1,394,592
• Trainable params: 1,394,216
• Non-trainable params: 376

CRNN - 4 classes (3 keywords + 1) - test acc: 96.23%

• Total params: 608,484
• Trainable params: 608,484
• Non-trainable params: 0

Autoencoder (+ SVM) - 4 classes (3 keywords + 1) - test acc: 92.36% - code dim: 12

• Total params: 1,531,381
• Trainable params: 1,531,381
• Non-trainable params: 0

PCA + SVM - 257 dim - test acc: 88.89%

• Trainable params: 66,049

PCA + SVM - 25 dim - test acc: 86.35%

• Trainable params: 625