An Improved Unvoiced Speech Segregation Based on Computational Auditory Scene Analysis

Monaural speech segregation based on computational auditory scene analysis (CASA) has been deeply developed in voiced speech segregation. However, it is much difficult to segregate unvoiced speech, due to its weak energy and lack of pitch period feature. Aiming at the uncertainty and unstability of background noise, an improved method for unvoiced speech segregation is proposed. After removing voiced speech segments, this method estimates the noise energy in each unvoiced segments by means of distance weighted noise estimation approach algorithm. Then, the spectral subtraction is utilized to extract and label the target unvoiced units. Compared with conventional unvoiced speech segregation method, the proposed method can improve the accuracy of residue noise estimation and attain better performance for unvoiced speech segregation.