Melody extraction for vocal polyphonic music based on bayesian framework

In order to automatically extract the main melody contours from polyphonic music especially vocal melody songs, we present an effective approach based on a Bayesian framework. According to various information from the music signals, we use a pitch evolution model describing how pitch contour changes and an acoustic model representing the acoustic characteristics when the pitch is a hypothesized one, and obtain the optimal melody contour utilizing a Viterbi algorithm. The experimental results on the RWC popular music database indicate that the overall accuracy achieves 73.4%.

Duke Scholars

Author Ming Li DKU Faculty