Beat-Synchronous Data-Driven Automatic Chord Labeling

Automatic Chord Labeling becomes a challenge when dealing with original audio recordings, in particular of modern popular music. In this work we therefore suggest a data-driven approach applying Support Vector Machines (SVM) and Hidden-Markov-Models (HMM) as opposed to typical chord-template modeling. The feature basis is formed by pitch-tuned chromatic feature information. For synchronization with the rhythmic structure we use IIR comb-filter banks for tempo detection, meter recognition, and on-beat tracking. The chord base is built by all typical triads resulting in 76 classes. A musiological model is used to model the context of a chord. Extensive experimental results are reported on 11k chords of 7h of MP3 compressed popular music and demonstrate effectiveness over the traditional approach to Automatic Chord Labeling: 60% accuracy are reached for this challenging task when major and minor chords are tagged.