A Unified System for Chord Transcription and Key Extraction Using Hidden Markov Models

A new approach for acoustic chord transcription and key extraction is presented. We use a novel method of acquiring a large set of labeled training data for automatic key/chord recognition from the raw audio without the enormously laborious process of manual annotation. To this end, we first perform harmonic analysis on symbolic data to extract the key information and the chord labels with precise segment boundaries. In parallel, we synthesize audio from the same symbolic data whose harmonic progression are in perfect alignment with the automatically generated annotations. We then estimate the model parameters directly from the labeled training data, and build 24 key-specific hidden Markov models for 24 different keys. The experimental results show that the proposed model not only successfully estimates the key, but also yields higher chord recognition accuracy than a universal, key-independent model.