Used ICPHI-2017 dataset to classify respiratory audio recordings to classify them on the presence of crackles and wheezles, which can be used by medical practitioners to diagnose respiratory diseases.
Ideas used: -Conversion audio files to spectrograms using librosa -Create CNN model to detect spatial features from spectrograms -Make prediction on basis of learned features for presence or absence of Crackle or Wheeze
Libraries/Frameworks: -Tensorflow -- to make the neural networks, do predictions. -Librosa -- to pre-process audio files -Numpy -- perform mathematical operations on image arrays -OpenCV -- to deal with images (here spectrograms) in python
Metrics obtained: -Accuracy -- 85.47% (test) -AUC score -- 0.76