Music Genre Classification using Machine Learning

Streaming services classify every uploaded track by genre, mood and instrumentation — and the classic academic version of that problem is GTZAN: 1,000 thirty-second clips, 10 genres, the benchmark Tzanetakis & Cook published in 2002 and hundreds of papers have used since. Raw audio is useless to a classifier directly; the work is in feature extraction — MFCCs that mimic human hearing, chromagrams for harmony, spectral contrast for timbre, tempo for groove — aggregated into a 58-dimensional fingerprint per clip. This project builds that full pipeline and trains a random-forest classifier on it, with honest stratified evaluation and a per-genre error analysis (rock vs metal is the classic confusion). The demo's spectrogram fingerprints and genre explorer make audio ML tangible. Built with standard tools (Python, librosa, scikit-learn).

1. The GTZAN collection (1,000 clips, 10 genres × 100, 30 s each at 22,050 Hz) is loaded and verified.
2. Each clip is framed (~1,300 overlapping windows); per-frame MFCCs, chroma, spectral features and zero-crossing rate are computed with librosa.
3. Frame features are aggregated (mean + variance) into a 58-dimensional clip fingerprint and standardized.
4. A stratified 80/20 split (artist-aware where possible) separates train and test; a random forest (300 trees) is trained.
5. The holdout is evaluated once: 84.2% accuracy is the design target; per-genre F1 and the confusion matrix go into the report.
6. The web demo serves the classifier behind the track picker with spectrogram fingerprints and probability bars.

Tech stack:

• Python 3, scikit-learn
• librosa (audio feature extraction)
• NumPy, pandas, Matplotlib
• Jupyter Notebook (training & evaluation)
• HTML5 canvas + JavaScript (demo)
• GTZAN genre collection (Tzanetakis & Cook, 2002)

Dataset & model details

• Dataset: GTZAN Genre Collection (Tzanetakis & Cook, 2002) — 1,000 thirty-second audio clips, 10 genres × 100 clips: blues, classical, country, disco, hiphop, jazz, metal, pop, reggae, rock; 22,050 Hz mono.
• Task: 10-class audio classification; input = 58-dimensional feature fingerprint per clip, output = genre probability distribution.
• Model: Random forest, 300 trees, on standardized MFCC/spectral features; CNN-on-mel-spectrogram documented as extension scope.
• Metrics: Accuracy and per-genre precision/recall/F1 on the stratified holdout — 84.2% accuracy is the design target for the built-to-order training run, stated honestly and never claimed as measured.

• [58-feature audio pipeline] librosa extraction: 13 MFCCs + deltas, 12-bin chromagram, spectral centroid/rolloff/contrast, zero-crossing rate, tempo — mean + variance aggregated per clip.
• [Random-forest genre classifier] 300 trees on the standardized fingerprint; a CNN-on-spectrogram variant documented as extension.
• [Track classifier demo] Pick from 10 demo tracks → predicted genre, confidence, mel-spectrogram fingerprint and top-5 probability bars.
• [Genre explorer] All 10 genres with their dataset fingerprints (tempo, timbre, energy) side by side.
• [Per-genre error analysis] Confusion matrix and F1 per genre — classical/jazz separate cleanly, rock/metal and disco/pop confuse, exactly as the literature reports.
• [Spectrogram visualizations] Mel-spectrogram-style fingerprints rendered per track in the demo.
• [Stratified honest evaluation] 80/20 split by clip with artist filtering where possible, so the model can't memorize performers.

• librosa feature-extraction pipeline (58 features per clip)
• Complete training & evaluation Jupyter notebook
• Trained random-forest genre classifier with inference code
• Interactive web demo (track classifier + genre explorer + spectrograms)
• Per-genre F1 and confusion-matrix analysis
• Project report PDF (audio-ML background, MFCC theory, methodology, honest evaluation)
• PPT presentation for final review
• Viva Q&A preparation document (MFCCs, spectrograms, random forests, stratified evaluation)

• 84.2% accuracy is a design target for the training run, stated honestly — the report documents the actual achieved figure after training.
• GTZAN is small and has known issues (a few mislabeled clips, artist repetition); the report discusses them and the evaluation mitigates with artist-aware splitting.
• Ten broad genres only — subgenres, mood and instrumentation tagging are out of scope.
• The demo uses illustrative track fingerprints; the delivered project trains on the real GTZAN clips.
• Real-world catalogs need far larger datasets and handle multi-label tracks; this is the clean academic version of the problem.