Classifying bird species from sound
Different acoustic representations reveal different structure; the model becomes stronger when each receives an appropriate processing branch.
Explore the multi-branch model.
A multi-branch classifier combining Mel, MFCC, and engineered acoustic features for imbalanced bird-species recognition.
Mel
3 × 100 × 60Three Mel representations retain time-frequency energy patterns. Each recording contributes 100 aligned time steps to this image-like branch.The heat map is an actual normalised 60-feature window from the executed notebook.
Every species gets equal weight.
Macro-F1 calculates a score per class before averaging, so frequent species cannot hide weak performance on rarer classes.
Classify bird species from recordings represented as 100 time steps and hundreds of acoustic features, while class imbalance makes raw accuracy misleading.
- 01
Explored feed-forward, convolutional, recurrent, and multi-input neural architectures.
- 02
Separated three Mel maps, nine MFCC maps, and 55 engineered features into specialized convolutional branches before fusion.
- 03
Used macro-averaged F1 so each species contributes equally to evaluation regardless of frequency.
The repository does not retain a distributable audio sample and runnable trained model, so this page explains the verified feature tensors and recorded evaluation without simulating a live prediction.
- PyTorch
- Python
- CNN
- MFCC
- Mel features
- Macro-F1