Multi-layer Attention Mechanism Based Speech Separation Model

Speech separation is the front-end of speech processing applications. Its purpose is to separate the speech in a multi-speaker environment. The neural network methods show good performance in speech separation, but most of the existing methods try to separate all the speaker speech. From the theory of auditory selection, we know that people can only focus on one speaker each time in multi-speaker conditions. Inspired by this, we use the attention mechanism to introduce the speaker information and propose a multi-layer structure so that the proposed model can extract a more complete separation speech. The experiments tested on the TSP and THCHS-30 datasets show that our model is superior to the baseline models in Short-Time Objective Intelligibility(STOI) and Perceptual Evaluation of Speech Quality(PESQ).