RespireNet

TensorFlow implementation of Microsoft's RespireNet to classify breathing/coughing sounds of people affected by COVID19 using the Coswara dataset. Once the train code is executed, the data is automatically downloaded and pre-precessed, thanks to the tensorflow-datasets pipeline. What you can customize:

• What audios to filter out: you can choose not to consider the wav files which last less than k seconds with k between 1 and 4, simply by passing skip=k when instantiating a CoswaraCovidDataset object
• Whether to use mixup: as a default, for the minority classes a kind of mixup augmentation is produced. Two examples of the same class are randomly selected and concatenated in order to create a new example. This helps in reducing class imbalance, but can be turned off siply passing mixup=False when instantiating a CoswaraCovidDataset
• What audio file you are interested in: the Coswara dataset is composed of 9 audiofiles per user, but my reduced version only considers four of them (breathing-deep, breathing-shallow, cough-heavy and cough-shallow). To select the one you are interested in simply pass audio_file=file_name when instantiating a CoswaraCovidDataset object
• The final lenght each audio file should have, by default this is set to 7 seconds.
• The low and high cuts for the 5th order Butterworth filetr applied on each data file to reduce noise
• The parameters used for the creation of the Mel Spectrogram using librosa

Requirements

The code requires Python >= 3.6 as well as the following libraries:

• matplotlib
• tensorflow_datasets
• pydub
• python-ffmpeg
• librosa
• scipy
• opencv_python
• tensorflow_gpu
• numpy
• tensorflow

Which can be installed easily through:

  pip install -U pip
  pip install -r requirements.txt

Finally, in order for the tfds audio processing pipeline you also need to have ffmpeg installed on your computer.

To-Do:

• Add a proper config file
• Find a way to filter out the audio files that only contain talking
• Add option to filter out asymtomatic people, as it doesn't make much sense to try and diagnose them using their cough or breathing