Separation of Speech from Speech Interference Based on EGG

Abstract

In this paper, a new method based on the Electroglottography (EGG) for solving speech separation problem was proposed. By using the EGG, the mixed speech signals was segmented to silence segment S, unvoiced segment U, and voiced segment V (SUV segmentation) according to the feature of the voiced and unvoiced speech. The V-segment algorithm was based on Blind Source Separation (BSS) using FastICA. Computational Auditory Scene Analysis (CASA) and spectral subtraction methods were used to obtain the target unvoiced Ideal Binary Mask (IBM). And the waveform synthesis was then performed to obtain the target unvoiced sound for U-segment. The experimental results showed that the proposed algorithm can improve SNR, similarity coefficient and subjective evaluation with EGG.