Audio-Classification

Problem Statement

Audio classification using Torch Audio and Librosa

Preprocessing

• All input audio files are processed to equal lengths using zero padding or cutting
• Audio downsampled to 22050 from 41000 to reduce complexity
• Mel Frequency Cepstral Coefficients (MFCC) used for collecting features
• MFCC converted the audio classification task to an image classification task

Training procedure

• Transfer learning using pretrained DenseNet for classifying the data. Resnet18, various CNN and RNN models were also tried out, but the results were unsatisfactory.
• Optimizer used is Adam with learning rate = 1e-3.
• Criterion used is Cross Entropy Loss
• Scheduler used StepLR whose parameters are from the PyTorch docs
• Training epochs: 20

Final results

• Training & validation accuracy: 98% (similar due to small train test split)
• Plots have been made towards the end of the report