Chord recognition using measures of fit, chord templates and filtering methods

This paper presents an efficient method for chord transcription of music signals. A succession of chroma vectors is calculated from the signal in order to extract the musical content of the piece over time. We introduce a set of chord templates for several types of chords (major, minor, dominant seventh, ...): different chord models taking into account one or more harmonics of the notes of the chord are considered. In order to fit the chroma vectors to the chord templates, we analytically calculate a scale parameter. The detected chord over a frame is the one minimizing a measure of fit between a rescaled chroma vector and the chord templates. Several popular measures in the probability and signal processing field are considered for our task. In order to take into account the time-persistence, we perform a post-processing filtering over the recognition criteria which quickly smooths the results and corrects random errors. The system is evaluated on the 13 Beatles albums and compared to the state-of-the-art. Results show that our method outperforms state-of-the-art methods but more importantly is significantly faster.