Research On Speech Separation Based On Deep Neural Networks

Speech is the most convenient and convenient form of human communication,and it is also the most natural way of human-computer interaction.However,in real life,the realistic background environment for both humans and automatic devices can be negatively impacted by background noise.Due to the large number applications of automatic speech recognition devices in daily life,separation of target speech from sound mixtures is of great importance.speech separation base monaural is most desirable from an application standpoint.The resulting monaural speech separation problem has been a central problem in speech processing for several decades.However,its success has been limited thus far.Because Deep Neural Networks(hereinafter referred to as DNN)has strong learning ability and is particularly suitable for studying the problem of speech separation,this paper focuses on how to use deep neural networks to improve speech perceptual quality and generalization performance.First,this paper finds a suitable set of speech separation features,and then uses this feature set to apply to speech separation based on DNN supervised learning.Largescale training can improve the generalization performance of speech separation.Then,for the improvement of speech perceptual quality,this paper studies a deep neural network architecture to directly reconstruct the time domain speech signal,and experiments and comparisons are made.The resulting system greatly improves the perceptual quality of speech.