Research For Self-attention Based Audio Source Separation Model

Audio source separation is an important and challenging signal processing problem.Audio source separation aims at separating target audio sources' signals from the mixture audio signal,which can be beneficial to the related signal processing works.Audio source separation has been researched for decades,but the related algorithms haven't completely solved this problem yet,so it is of great significance to do research work on it.With the advent of the big data and the improvement of hardware's computing ability,deep learning has shown excellent performance among various fields of computer science.Hence,researchers have begun to attempt to use deep learning to solve problems which are difficult to be solved with traditional algorithms.This thesis chooses to use convolutional neural networks(CNN)to solve the audio source separation problems to complete the audio source separation tasks.The works of the thesis are mainly as follows:First,the thesis has analyzed the problem of traditional CNN which could lose important global information during the calculation process due to the characteristics of the local receptive field of the convolution itself,and further proposed to introduce the self-attention mechanism into CNN to restore the losing information by calculating the self-attention feature maps.Further,the thesis also synthesizes recent related works and proposes a dual-path simplified self-attention mechanism,which can not only reduce the parameters of the traditional selfattention mechanism but also improve the performance of the self-attention model.Second,based on the work of the first point above,the thesis has further analyzed the problem of information losing around the band borders caused by the band splitting.Hence,the thesis has proposed a new and general cross-band features to solve the problem and improve the existing methods.Last,the thesis has designed an application for the accompaniment extraction tasks,whose model has been trained and tested on the public datasets.By analyzing the experimental results,the validity and practicability of the proposed audio source separation model have been verified.