Abstract
This thesis proposes an audio melody extraction based on hidden Markov model (HMM) to extract the singing pitches from polyphonic music. The goal is to improve the raw pitch accuracy by applying different methods to select more stable peaks. First, fast Fourier transform is applied to divide a reasonable frequency range of human voice into several frequency bins, and these bins are considered as different states in the HMM. Second, we take advantage of the temporal and spectral variability between human voice and instruments to filter out most of the background instruments from the origin signal. After that, we apply several methods, such as normalized sub-harmonic summation, to select the peaks at each frame as features, and each state is modeled as a GMM. Lastly, Viterbi decoding is performed based on the transition probability between states and state occupation likelihood to find out a continuous singing pitch contour. Experimental result shows that the proposed approach achieves a better performance than the single peak extraction approach.