Speech Enhancement with Phase Correction based on Modified DNN Architecture

Speech enhancement is an important issue in the field of speech signal processing. With the development of deep learning, speech enhancement technology combined with neural network has provided a more diverse solution for this field. In this paper, we present a new approach to enhance the noisy speech, which is recorded by a single channel. We propose a phase correction method, which is based on the joint optimization of clean speech and noise by deep neural network (DNN). In this method, the ideal ratio masking (IRM) is employed to estimate the clean speech and noise, and the phase correction is combined to get the final clean speech. Experiments are conducted by using TIMIT corpus combined with four types of noises at three different signal to noise ratio (SNR) levels. The results show that the proposed method has a significant improvement over the referenced DNN-based enhancement method for both objective evaluation criterion and subjective evaluation criterion.