Acoustic signal classification

01 / CONTEXT

The challenge.

Environmental sounds vary in duration, frequency content, and recording conditions. The project investigates which representations and models can separate sound categories effectively.

02 / PROCESS

The approach.

1. Use ESC-50: 2,000 environmental recordings across 50 categories.
2. Compare STFT, MFCC, and Log-Mel feature pipelines using Librosa and PyTorch.
3. Compare an augmented four-block CNN with a fine-tuned ResNet-50 model.

03 / REFLECTION

The outcome.

Compared MFCC features, a custom CNN, and ResNet-50 on environmental sound classification. Reported results were 44.1% for the MFCC baseline, 74.25% for the custom CNN, and 81.25% validation accuracy for ResNet-50.