From Sound to Label a Music Classification System with Wav2Vec2 Base 960h

This work explores the genre classification of music using the Facebook Wav2Vec2-Base-960h model, a self-supervised deep learning model originally designed for speech representation. Raw audio waveforms are transformed into informative feature representations by fine-tuning the model on heterogenous song datasets, which are then passed through a classifier to be labeled by genre.This process bypasses the laborious hand-crafted feature extraction process by taking advantage of Wav2Vec2’s highly efficient audio analysis. Experimental results demonstrate competitive classification accuracy, which indicates the efficacy of self-supervised learning. The proposed framework offers a scalable and efficient solution for music streaming services and digital libraries.