Portfolio
AI / Machine Learning|2024

Music Genre Classification

End-to-end reproducible pipeline achieving SOTA 83.5% accuracy on GTZAN with a U-Net inspired model. Leak-free methodology with track-level splits, cross-validation, and transfer learning.

PythonPyTorchScikit-learnJupyterMel-Spectrograms

Classification Confusion Matrix

Rows: actual genre, Columns: predicted genre

Blues
Classical
Country
Disco
HipHop
Jazz
Metal
Pop
Reggae
Rock
Blues
82
2
4
1
0
5
0
2
3
1
Classical
1
95
0
0
0
2
1
1
0
0
Country
3
0
78
2
1
1
0
5
4
6
Disco
1
0
2
80
5
0
2
4
3
3
HipHop
0
0
1
4
85
1
3
2
2
2
Jazz
4
3
1
0
0
88
0
1
2
1
Metal
0
1
0
2
2
0
90
1
1
3
Pop
1
1
4
5
2
1
1
78
3
4
Reggae
2
0
3
3
2
2
1
3
80
4
Rock
2
0
5
4
3
1
4
4
3
74

Motivation

Music genre classification sounds like a solved problem, until you look at how most papers evaluate their models. The standard GTZAN benchmark has a well-known data leakage issue: audio tracks are sliced into segments before splitting into train/test sets, meaning segments from the same song can appear in both. This inflates accuracy numbers and makes results unreproducible.

I wanted to build a pipeline that does it right: rigorous methodology first, state-of-the-art accuracy second.

The Approach

Leak-free data pipeline: the critical step is splitting at the track level (60/20/20) before any audio slicing. Each 30-second track is sliced into 10 segments of 3 seconds only after the split. This guarantees no information leakage between train and test sets. This distinction changes reported accuracy by 5-10% compared to naive approaches.

Feature extraction: 128-bin log Mel-spectrograms computed from each 3-second segment, with the scaler fitted exclusively on training data. This is another common source of leakage that many pipelines miss.

Architecture tournament: three CNN architectures were designed and compared:

  • Efficient_VGG: lightweight baseline inspired by VGG with reduced parameters
  • ResSE_AudioCNN: residual blocks with squeeze-and-excitation attention
  • UNet_Audio_Classifier: an encoder from U-Net architecture repurposed for classification

Results

The U-Net encoder architecture achieved the best performance:

  • 82-83% test accuracy on GTZAN with proper leak-free evaluation
  • ~90% cross-validation mean demonstrating consistent performance
  • Strong transfer to Indian Classical Music and Tabla Taala datasets without fine-tuning

These numbers are lower than many published results on GTZAN, and that's by design. Papers reporting 90%+ typically have data leakage in their evaluation pipeline. Our 83% with proper methodology is a more honest benchmark.

What I Learned

The biggest takeaway was that evaluation methodology is as important as model architecture. The U-Net encoder wasn't a novel idea, but combined with a rigorous, leak-free pipeline, it outperformed supposedly superior architectures that were evaluated with flawed methodology. In ML research, honest evaluation is itself a contribution.

This project was a collaboration with Camilla Sed.