mlachmish/MusicGenreClassification

# Technical Summary Applies convolutional neural networks to mel-frequency spectrograms rather than raw audio or traditional MFCCs, processing 599×128 feature vectors through three convolutional layers with max pooling before softmax classification. Built with TensorFlow and trained on a custom dataset of ~10,000 music previews sourced from the Million Song Dataset via the 7Digital API, addressing the limited scale of prior academic work by tackling all 10 genre classes simultaneously. Uses librosa for audio preprocessing with 100ms windows and 40ms stride to extract mel-frequency features, demonstrating improved accuracy over previous approaches that relied on smaller datasets or RBM architectures.

598 stars. No commits in the last 6 months.