A Multipitch Tracking Algorithm Based on Wavelet Packet Analysis

Abstract

In view of the limitation of many multipitch detection methods in single-channel speech, a robust and accurate multipitch estimation method for multiple voices is processed. In this study, our approach is based on the spectral analysis of the wavelet packet analysis. It utilizes the quasi-periodicity in a short time frame of speech, and gets candidate pitche in every frame by using peak selection and searching algorithm. The multipitch envelope of the mixed signal is obtained by neighborhood analysis and determined from these candidates by virtue of the fact that pitch should change rather smoothly in consecutive frame. Simulation results showed that the proposed method can robustly estimate fundamental frequency for mixed speech from single-channel speech.