Case Study
GenreAI
A neural-network music-genre classifier trained on engineered audio features.
Jun 2024
The problem
Can a lightweight neural network tell musical genres apart from audio alone? GenreAI was my deep dive into audio ML - building a genre classifier from feature extraction up, rather than throwing a large model at raw waveforms.
The approach
librosa extracts a comprehensive feature set from each track: MFCCs, chroma STFT, RMS, spectral features, zero-crossing rate, harmony and tempo. A TensorFlow/Keras neural network trains on these features over the GTZAN benchmark dataset - 1,000 audio tracks across 10 genres, from blues and classical to jazz and rock.
The outcome
After extensive experimentation and hyperparameter tuning, the model reached 99.67% classification accuracy. The entire project is open source on GitHub - code, trained model and replication instructions included.
GTZAN Benchmark Comparison
GenreAI - 99.67%
GenreAI achieves 99.67% classification accuracy on the GTZAN 10-genre benchmark, approaching the highest published results while outperforming representative classical machine learning and standard CNN baselines.
Learnings
Feature engineering did the heavy lifting: the performance gains came from richer audio-feature extraction and systematic tuning, not from a bigger network.