Nepali Music Genre Classification Using CNN-SVM Hybrid Architecture

Abstract

Nepali music genre classification using CNN–SVM hybrid model was developed to address the challenge of categorizing local genres such as Gazal, Lok Dohori, Nephop, and Pop. A dataset of 1,000 manually curated songs (250 per genre) was collected from YouTube, segmented into 30‑second clips at 25%, 50%, and 75% of each track’s duration, resulting in approximately 3,000 audio segments. Each segment was converted into a 128×128 Log‑Mel spectrogram. A four‑layer CNN extracted a 64‑dimensional embedding, which was then passed to an SVM classifier. Experiments showed that the CNN–SVM hybrid with an RBF kernel achieved 88.29% accuracy, outperforming the standalone CNN baseline (84.28%). Among evaluated kernels, RBF and Linear both achieved the highest accuracy of 88.29%, while the Sigmoid kernel performed worst at 79.60%. The results demonstrate that combining deep learning feature extraction with a traditional machine learning classifier is effective for Nepali music genre classification on moderate‑sized, domain‑specific datasets.