Research On Speech Separation Algorithm Based On Deep Neural Networks

In real environment,people's target speech signal is usually disturbed by noise,which will seriously damage the intelligibility of speech and reduce the performance of speech recognition system.For noise interference,front-end speech separation technology is one of the most commonly used methods.A good front-end speech separation system can greatly improve the intelligibility of speech and the recognition performance of automatic speech recognition system.Speech separation has shown good prospects in many fields,but there are also a lot of shortcomings and challenges.The complexity of speech signal makes the acoustic characteristics of the same kind of speech signal very different,which puts forward higher requirements for the learning ability and generalization ability of speech separation system.In addition,the overlap of speech signals in real environment makes it difficult to extract target characteristics.Therefore,compared with traditional audio processing technologies such as speech recognition,there is still a gap in the separation performance of speech signals.Speech separation is a technology for separating target speech from mixed speech.Speech separation involves a wide range of applications,including voice calls,teleconferencing,scene recording,military eavesdropping,speech recognition system,hearing aids and speech recognition equipment.In order to improve the performance of speech separation system,this paper focuses on three aspects: network design,feature selection and learning objectives.The main work and contributions are as follows:(1)In the aspect of network framework design,we study the principles and methods of the traditional pattern recognition classifier,and realize the speech separation based on the traditional classifier.Understanding the principles and methods of neural network,combining the convolution neural network and the fully connected network,designing the deep stacked residual network,taking the traditional and spectral information as the input of the network,has a certain improvement in the speech intelligibility and voice quality compared with the existing algorithms.(2)In the aspect of feature selection,we study the principles and methods of the fireworks algorithm,apply the fireworks algorithm to feature selection,and propose an improved speech separation model for feature selection problems in traditional speech separation,and feature selection is added to ensure the validity of the extracted feature.(3)In terms of learning objectives,a hybrid masking method based on ideal binary masking and ideal floating-value masking is proposed.Experiments show that the speech quality of PESQ is improved by 1.6% compared with ideal floating-value masking.In this thesis,full comparative experiments have been done on the above three works.The experimental results show that these methods have greatly improved the speech quality and intelligibility.