Research · Under review · TISMIR
African Music Genre Classifier (AFMGC)
- Data
- 1,061 human-annotated 30-second clips, ten genres
- Features
- 91 acoustic descriptors per clip plus aligned Mel-spectrograms
- Models
- Classical and convolutional baselines
- Context
- PGD thesis, AIRLAB, University of Lagos
01 — Motivation
Most music-information-retrieval corpora are built around Western repertoire, so models trained on them flatten a continent of music into “world”. AFMGC is an annotated corpus of African popular music and a deep-learning pipeline for genre classification, labelled by listener judgement rather than inherited metadata.
02 — Data
1,061 human-annotated 30-second clips across ten West, Southern and Central African genres, balanced by class and standardised to mono, 22,050 Hz, uncompressed. Each clip is annotated in Label Studio for genre, lyrical language and instrumentation by listener judgement.
03 — Method
A reproducible Python pipeline (Librosa, NumPy): silence trimming, loudness normalisation and STFT framing, then MFCC, spectral and chroma descriptors combined into a 91-dimensional representation, with aligned Mel-spectrograms for deep models. Classical and convolutional baselines give future users a reference point.
04 — Analysis
Correlation analysis, PCA and t-SNE show where genres separate in feature space and where historically related genres overlap.
05 — What’s next
Extending the corpus toward self-supervised pre-training and low-resource transfer learning, and testing whether music representations pre-trained on Western repertoire transfer to African genres at all.