Large-vocabulary Chord Transcription Via Chord Structure Decomposition

While audio chord recognition systems have acquired considerable accuracy on small vocabularies (e.g., major/minor chords), the large-vocabulary chord recognition problem still remains unsolved. This problem hinders the practical usages of audio recognition systems. The difficulty mainly lies in the intrinsic long-tail distribution of chord qualities, and most chord qualities have too few samples for model training. In this paper, we propose a new model for audio chord recognition under a huge chord vocabulary. The core concept is to decompose any chord label into a set of musically meaningful components (e.g., triad, bass, seventh), each with a much smaller vocabulary compared to the size of the overall chord vocabulary. A multitask classifier is then trained to recognize all the components given the audio feature, and then labels of individual components are reassembled to form the final chord label. Experiments show that the proposed system not only achieves state-of-the-art results on traditional evaluation metrics but also performs well on a large vocabulary.