Speech Segregation based on Pitch Track Correction and Music-Speech Classification

A novel approach for pitch track correction and music-speech classification is proposed in order to improve the performance of the speech segregation system. The proposed pitch track correction method adjusts unreliable pitch estimates from adjacent reliable pitch streaks, in contrast to the previous approach using a single pitch streak which is the longest among the reliable pitch streaks in a sentence. The proposed music and speech classification method finds continuous pitch streaks of the mixture, and labels each streak as music-dominant or speech-dominant based on the observation that music pitch seldom changes in a short-time period whereas speech pitch fluctuates a lot. The speech segregation results for mixtures of speech and various competing sound sources demonstrated that the proposed methods are superior to the conventional method, especially for mixtures of speech and music signals.