Speech Enhancement Based On Deep Learning

Speech,as the acoustic representation of language,is one of the most natural,effective,and convenient ways of human communication.Speech applications such as smart headset,cell phone,and voice assistant have been widely used.In real world,people or machines are inevitably interfered by various noises while capturing speech signals.These interferences corrupt the received signal.The speech captured in practice is mainly interfered by three components: environmental noise,room reverberation,and other interfering speakers.The goal of speech enhancement is to overcome these three obstacles.Traditional algorithms are based on statistics which are difficult to handle the challenges of real-world scenarios.In the past few years,deep learning has developed rapidly.Compared with traditional algorithms,deep learning-based methods convert the speech enhancement problem to a machine learning problem.The methods train a data-driven model by a large number of data leading to significant improvements.In this dissertation,we study deep learning-based speech enhancement methods to explore ways to combine speech signal processing knowledge with deep learning.Firstly,a signal-to-noise ratio(SNR)estimation method is proposed.The SNR reflects the degree of speech interference by noise,which is an important parameter,and accurate estimation of the SNR can contribute to speech enhancement tasks.We also investigate the tasks of noise reduction,reverberate cancellation,and target speaker separation.The content and innovations of this paper are as follows:(1)SNR estimation.18 common acoustic features are analyzed,respectively.Feature combinations is also systematically investigate using Group Least Absolute Shrinkage and Selection Operator(Group Lasso)and sequential floating forward selection(SFFS)algorithms.The results show that the feature combination can further improve the performance and obtains a higher accuracy SNR estimation.(2)Noise reduction.For the ambient noise scenario,a dual-microphone based speech enhancement model by combining beamforming and deep learning is proposed.By analyzing different input features of dual-microphone speech enhancement,we find that the features based on two differential beamformers with opposite directions have the advantages of easy calculation,reflecting spectral information and frequency invariance.Especially in the case where microphones are placed close,the frequency invariant characteristics of the differential array can convert the various phasedifference between different frequency bands into uniform amplitude-difference.Finally,the extracted features are used as the input of the deep neural network(DNN).The performance is significantly improved compared with the baseline system.(3)Reverberation Cancellation.For the reverberant scenes,a noise-robust reverberation cancellation algorithm is proposed,which combines Weighted Prediction Error(WPE)in traditional signal processing and deep learning.Meanwhile,considering the noise in real environment,speech and noise have different echo paths,and the sparse characteristics of speech,a dual filter strategy is proposed for reverberation cancellation,which achieves a noise-robust reverberation cancellation performance.(4)Target speaker separation.For the speaker interference scenarios,a target speaker separation algorithm based on a dynamic attention mechanism is proposed.By analyzing the target speaker separation problem using Anchor speech,Dynamic Attention is introduced to efficiently capture the target speaker information contained in the Anchor speech.The mechanism improves the performance of target speaker separation in the Encoder-Decoder framework.