Abefocuson / OSF-Audio-Classification

Machine learning models for OSF cough audio data classification

OSF cough audio dataset classification

This repository provides several machine learning models for OSF cough audio data classification using Keras with tensorflow backend. The dataset can be found here.

Models:

1. Using MFCC arrays as input
   MFCC_OSF.ipynb (61% accuracy)

2. Transfer learning with pretrained VGG16 network based on ImageNet
   TransferLearning_OSF_Melspectrograms.ipynb (83% accuracy)
   TransferLearning_OSF_spectrograms.ipynb (80% accuracy)
   TransferLearning_OSF_wavelet.ipynb (85% accuracy)

3. Transfer learning with pretrained VGGish network based on Google AudioSet
   VGGish-TransferLearning_OSF_Melspectrograms.ipynb (74% accuracy)

Due to timing I didn't do much tuning on hyperparameters. Based on the current results, transfer learning using pretrained VGG16 network has the best performance in terms of accuracy and confusion matrix, although it takes the longest training time. I believe all the models would perform better given more time for tuning.

Reference:

• Transfer Learning with VGG16

• VGGish

• VGGish to Keras