Dataset, Features, Model, and Quantization Strategy for Respiratory Sound Classification

Mel-spectrograms feed a CNN-RNN; last layers retrain on tiny patient sets, then log-quant weights trim memory 4× for wearables.