Victor Simonin
Alexandre Lemonnier
The effective diagnosis of COVID could have been an effective tool for limiting coronavirus transmission if our public policies were effective and had implemented the test-trace-and-isolate solution.
Unfortunately, COVID tests require individuals to go to specific locations for testing and take time.
The goal of this project is to develop a sound classification model capable of distinguishing a COVID-related cough from a benign cough using recordings of a patient's cough.
pip install -r requirements.txtTo predict if a cough is indicative of a positive or negative COVID case from an audio file, simply run the following command:
python predict.py <audio_path>The classification model covid_cough_classifier.h5 was generated in the covid_audio_classification.ipynb notebook. In this notebook, the librosa library was used to extract the mel-spectrograms from the audio files in the Coswara-Data dataset.
The keras library was then used to build a convolutional layer classification model to predict if a cough is indicative of a COVID-19 case or not. However, the model was not trained on a dataset of sufficient size to predict with high accuracy, and overfitting is present.
Despite this, the model is still able to predict with 70% accuracy if a cough is indicative of a COVID-19 case or not on a test dataset.
The overfitting may be due to the lack of significant distinguishing elements in the audio files of the coughs that would allow for COVID-19 prediction in the Coswara-Data dataset, and the size of the dataset used to train the model, which is in the following form:
- Train data : negative (1089) | positive (495)
- Validation data : negative (273) | positive (124)
- Test data : negative (341) | positive (155)