Music Genre Classification

This review documents a reproducible pipeline for automatic Indian music genre recognition that converts short (≈30 s) audio clips into spectral features and evaluates multiple classical classifiers. The implemented workflow covers dataset organization, FFT‑based feature extraction (the first 2000 frequency bins saved as reusable .npy files), model training with SVM, MLP, KNN and logistic regression, cross‑validation evaluation, confusion‑matrix diagnostics, and a single‑file tester for qualitative checks. Key findings are that FFT magnitudes provide a simple, interpretable baseline enabling working classifiers to separate several genres reliably; that support vector machines and carefully tuned MLPs generally outperform simpler models on these high‑dimensional spectral vectors though overall performance remains constrained by the chosen features and dataset quality; and that common failure modes are consistent confusions between acoustically similar genres, which exposes the fundamental limitation of global FFT representations that discard temporal dynamics. The review also notes practical reproducibility issues arising from hard‑coded paths, deprecated imports, and missing environment manifests. To address these gaps, it recommends moving to perceptual, time‑aware features such as mel‑spectrograms or MFCCs, applying scaling and PCA, adopting stratified hold‑out testing and principled data augmentation (e.g., SpecAugment, mild time/pitch perturbations), and supplying a requirements file with relative model paths. Overall, the project establishes a transparent, low‑compute baseline useful for comparative research, cultural archiving, and metadata enrichment and provides a clear roadmap toward spectrogram‑based and pretrained deep‑learning approaches for improved performance.