Health

Respiratory Sound Classifier

A model that listens to a few seconds of breathing or coughing and predicts which of five respiratory conditions it sounds like. It reuses Google's YAMNet audio model to understand the sound and trains a small network on top.

Gets 87% of held-out clips right across five classes while training a network of only about 68 thousand parameters.

87% of held-out clips classified correctly across five conditions

Each condition leaves a different pattern in the sound

COPD and healthy lungs are the easiest to tell apart

01 / 0387% of held-out clips classified correctly across five conditions

Overview

Doctors learn a lot from how someone breathes, but a stethoscope is only useful when a trained ear is listening. This project explores whether a model can pick up the same patterns from a short recording, as a starting point for low-cost screening tools.

It uses the balanced ICBHI respiratory sound dataset with five labels: Bronchial, asthma, COPD, healthy and pneumonia. Each clip goes through YAMNet, a model Google trained on millions of everyday sounds, which turns it into 1024 numbers that describe what it hears. A small PyTorch network then learns to map those numbers to a condition.

The result is a research prototype, not a diagnostic device. A Colab notebook walks through the whole pipeline: loading the data, extracting features, training, and plotting how well it does.

Features

Classify a breathing or cough recording into one of five respiratory conditions
See confidence scores for every class, not just the top guess
Retrain the classifier end to end with one script or the Colab notebook
Inspect results with a confusion matrix, per-class report and learning curves
Run predictions on new audio through a small inference module

Challenges

The dataset is small, so a large model memorizes it. Training on frozen YAMNet embeddings with a compact network, batch normalization and 60% dropout kept validation accuracy close to training accuracy.
Classes are uneven, so the loss uses balanced class weights and label smoothing, and the train and validation split is stratified so every condition shows up in both.
Recordings vary in length, while the model expects a fixed window. Every clip is resampled to 16 kHz and padded or trimmed to the same 4.35 seconds used in training, both when training and at prediction time.
An earlier version used PANNs features at 4096 dimensions. Switching to YAMNet's 1024-dimensional embeddings made the model smaller and less prone to overfitting.