Research On Separation Of Audio Signal Based On Deep Neural Network

As a kind of art,music has a long history of development and has produced many popular works of art.With the improvement of computer performance,audio processing technology has also made great development.In recent years,deep learning technology has been used to study the separation of audio signals.Has become an increasingly popular topic in the area of audio signal processing.Promote the development of audio separation based on deep learning technology.The audio separation model based on the deep neural network(5Layer-DNN model)is used to separate the speech from the speech.The Perceived Speech Quality Evaluation(PESQ)is introduced and used to evaluate the separation results.Analyzing PESQ evaluation results of original mixed audio,separated speech and piano audio under different signal-to-noise ratios,it is found that the speech component and the piano component obtained by the audio separation of the mixed speech with the SNR of 5dB are the best.By comparing 40,50,60 training results of three different PESQ evaluation results show that when the number of iterations is 50,the model has converged;5Layer-DNN model and L-MMSE algorithm comparison shows that this paper uses 5Layer-DNN model in speech and audio separation more performance excellent.The 5Layer-DNN model was used to separate the mixed speech composed of voice and piano accompaniment and the PESQ index was used to evaluate the separation results.By comparing the PESQ evaluation results of 40,50,and 60 iteration times,it is shown that the model converges when the number of iterations is 40;comparing the 5Layer-DNN model with the L-MMSE algorithm shows that the 5Layer-DNN model in speech source Better separation performance.The 5Layer-DNN model,the 6Layer-DNN model and the CNN model are respectively used to separate the audio source of the mixed music of piano and violin.The results show that the 5Layer-DNN model is more effective in separating the piano components of the mixed instrument audio,and the 6Layer-DNN model is more effective in separating the violin components.The separation effect of CNN model is worse than the former two.The average SIR of the piano and violin components obtained from the three instrument models was 9.6 and the average subjective MOS score was between 3 and 3.5(5 out of full)indicating the three models used in this study,especially The 5Layer-DNN model is effective at separating mixed-source sounds...