Single Channel Speech Separation Methods Based On Deep Neural Network

Speech separation mainly includes the separation of clean sound and clean sound,the separation of noise and noise,and the separation of clean sound and noise.The main work of this paper is to study the separation of clean sound and noise.With the maturity of artificial intelligence technology,there are more and more applications relying on voice interaction technology in real life.However,some noise around them often interferes with the progress of voice interaction and reduces the performance of voice interaction.Therefore,voice separation technology becomes particularly important.For noise interference,one of the most common solutions is to use front-end speech separation technology.However,the voice separation technology in real environment is far from ideal.Especially in the case of unstable noise,voice separation technology is still facing severe challenges.In recent years,with various deep-seated models being excavated continuously,the research of speech separation technology based on deep neural network(DNN)has become extremely popular.Because DNN has strong non-linear modeling ability compared with other models,it can better separate speech,and it becomes a trend to use DNN in the field of speech separation.The main works of this paper are as follows:(1)The depth of this article is based on the neural network research and implementation of speech separation method,In view of the shortcomings of existing DNN models,a combination of multiple models(CE-DNN)speech separation method is proposed.Two different training sets are put into DNN for training,and two different parameters of DNN training models are obtained.Then the test data are put into two training models and the output results are combined.The pure speech is mixed with Different types of noise,and the experiments are carried out with different input signal-to-noise ratios of noise.The experimental results show that compared with the original DNN model,CE-DNN model cannot only improve HIT-FA(hit rate-false alarm rate)in ideal binary masking,but also improve short-term objective speech intelligibility of speech targets.(2)This paper also proposes a speech separation model(WF-DNN),which combines DNN with Wiener filtering,namely to extract speech sound level characteristics,learning through within DNN to separate target voice map with noise characteristics,separation target speech,then using wiener filtering method for separation of target speech make a noise reduction processing,after the wiener filter is obtained by inverse transformation of the separation of speech waveform signal.In this paper,two experiments based on IBM training target and IRM training target were completed.Experimental results show that compared with the original DNN model,the WF-DNN model can effectively remove noise interference and improve the signal-to-noise ratio of separated speech.