Accurate Labeling of Time-Frequency Units in Monaural Voiced Speech Segregation

Segregation of speech in presence of noise and other interferences is a challenging problem, especially in monaural case. Since accurate labeling of time-frequency (T-F) units has great impact on segregation results, in this paper we propose a new method for labeling T-F units in monaural voiced speech segregation which marks these units using envelope cross-channel correlation in the segmentation stage and label them with different features in low and high frequencies in the grouping stage. We use enhanced envelope autocorrelation function (EEACF) for labeling T-F units in the high frequency range. The evaluation of the system including subjective and objective criteria is done for different types of noise mixed with speech signals which shows better segregation results than conventional methods.