Abstract:

Speech separation is the process of separating the target speech and noise from the noisy speech mixture. Speech separation algorithms are useful in improving the quality and intelligibility of the speech. The various traditional speech separation algorithms such as spectral-subtractive algorithms, Wiener filtering, statistical model-based methods and subspace algorithms are mainly focus on improving the speech quality. But there are applications such as mobile communication, air ground communication and hearing aids, needs speech intelligibility than speech quality. In order to satisfy the requirements of intelligibility, this work proposes an algorithm using Computational Auditory Scene Analysis (CASA) and Support Vector Machine (SVM) to separate the noisy speech into target speech and noise and at the same time improves the speech intelligibility. The proposed algorithm decomposes the clean speech and noise into time-frequency units (T-F) and computes the energy from each frame of clean speech and noise to train the SVM. Latter in the testing phase, the trained SVM is used to estimate the binary mask from the energy of the noisy speech based on whether each T-F unit is dominated by speech or noise. The estimated mask by SVM is used to synthesize the speech signal and is presented to normal-hearing listeners with different age groups to measure the performance of the proposed algorithm. The experimental results show substantial improvements in recognition score because the separated speech has reasonable speech intelligibility.