What it does

Trains a convolutional neural network to classify 10-second audio clips into one of ten genres (blues, classical, country, disco, hiphop, jazz, metal, pop, reggae, rock). The pipeline downloads 30-second song previews via the 7Digital API, converts them to mel-frequency spectrograms using librosa, and feeds a 3-layer CNN with max pooling and softmax output.

The interesting bit

The author couldn’t find a clean labeled dataset, so he reverse-engineered one: the Million Song Dataset provides metadata and 7Digital IDs, and 7Digital’s “preview before you buy” feature becomes a free data spigot. He also discovered that raw mel-frequencies (step 2 of the MFCC pipeline) outperform full MFCCs by “extremely better” margins, at the cost of longer training — a finding backed by t-SNE visualizations showing cleaner genre clustering.

Key highlights

• Built on TensorFlow in 2016, explicitly framed as a learning exercise for the then-new framework
• CNN architecture with 3 hidden layers and max pooling, inspired by Sander Dieleman’s Spotify blog post on deep content-based recommendation
• Dataset construction via previewDownloader.py scraping 7Digital previews for ~1% of the Million Song Dataset (roughly 2.8GB due to laptop constraints)
• Preprocessing scripts for MFCC and mel-spectrogram extraction, t-SNE visualization, and input formatting all included
• Results published for full 10-class classification, compared against Tao Feng’s 2/3/4-class RBM results and a non-deep-learning benchmark

Caveats

• The README contains no actual accuracy numbers in text; results are only visible in the results_mine.png image, so precise performance is unclear without viewing it
• Dataset download link is a Dropbox URL of uncertain longevity
• Spelling inconsistencies (“nural_network.png”, “GenereClassification”) suggest limited maintenance since original publication

Verdict

Worth a look if you’re teaching or learning classic audio CNN pipelines, or if you need a reference for scraping creative datasets from commercial APIs. Skip it if you want a maintained, production-ready classifier — this is academic coursework from the TensorFlow 0.x era, not a library.