Signal Processing & ML

Music ML — Playlist Classifier

Multi-label song classifier predicting playlist assignments across 11 genre and mood categories. Built on Sentence-BERT embeddings and a multi-output MLP — and preceded by a full OCR-based data recovery pipeline that reconstructed an 18,000-song library corrupted during platform migration.

GitHub

Scale

18K

Songs Recovered via OCR

Playlist Labels

Pipeline

OCR Extraction

Tesseract OCR over screenshots of the music library UI — extracts raw title/artist text per page image.

OCR Cleaning & MusicBrainz Validation

Regex cleaning removes timestamps and UI fragments. Adjacent short lines are paired as Title–Artist candidates and validated against MusicBrainz (rate-limited at 1 req/sec).

Metadata Enrichment

Each song enriched with Genius lyrics, Last.fm genre tags, NLTK VADER sentiment, and NRCLex dominant emotion. Resume-safe script skips already-processed rows.

Classification

Lyrics + tags + sentiment + emotion concatenated and encoded to 384-dim vectors via Sentence-BERT (all-MiniLM-L6-v2). MLPClassifier(256, 128) wrapped in MultiOutputClassifier — one binary classifier per playlist. Per-class probability thresholds (0.15–0.50) tuned to balance recall vs. precision per genre.

Model Details

Component	Details
Embedding	`all-MiniLM-L6-v2` — 384-dimensional Sentence-BERT vectors
Classifier	`MLPClassifier(hidden_layer_sizes=(256, 128), max_iter=300)`
Wrapper	`MultiOutputClassifier` — independent binary classifier per label
Class Balancing	Upsample with replacement to max class size before training
Runtime	~10 min on Colab T4 GPU for 241K songs

Playlists

rapGenre

countryGenre

popGenre

edm / alt rockGenre

christianTheme

christmasSeasonal

lofiMood / Style

feelsMood

movieContext

starCurated

vocalist / instrumentalStyle