Mirex 2010: Joint Recognition of Key and Chord from Music Audio Signals Using Key-modulation Hmm

This extended abstract describes a submission to the Music Information Retrieval Evaluation eXchange 2010 (MIREX 2010) in the Audio Chord Estimation and Audio Key Detection tasks. We propose a new model to recognize musical keys and chords simultaneously from musical acoustic signals including key modulations. Chords and keys are closely related notions of music involving harmony. Since occurrences of the chords and transitions between chords largely depend on its keys, it is obvious that using key information supports chord recognition task. But the problem here is that keys and chords are both unknown and the possibility of key modulations within a song makes this problem even harder. Considering the mutual dependency of keys and chords, we believe that recognizing both of them at the same time helps one another than recognizing them separately. To realize this, we propose keymodulation hidden Markov model (HMM) whose hidden states representing each pair of key and chord. The model has transition probabilities to different keys which can handle key modulations. By applying the Viterbi decoding, the key and chord for each frame is estimated. The experimental result shows that this model performs better for chord recognition than the conventional chord HMM and can also recognize keys in high accuracy.