Research On Robust Speech Separation Method

Speech separation technology is one of the research hotspots in speech signal processing.Because a single microphone is not only easy to deploy,but also relatively low cost.So the single channel speech separation technology has been widely used in the scene and a certain practical significance.Single-channel speech separation requires the recovery of multiple speech signals from one-dimensional mixed signals,which is an underdetermined problem and very challenging to solve.Moreover,in the real scene,the desired target signal will not only be interfered by the voice of others,but also be polluted by environmental noise,which greatly increases the difficulty of speech model separation.Aiming at the problem of single channel speech separation with background noise,this paper proposes a speech separation method based on uncertainty perception and a dual link speech separation method combined with elastic weight consolidation strategy.Experiments on open data sets show that the proposed method can improve the separation quality of speech models.The main work contents and contributions of this paper include the following two aspects:(1)Speech separation based on uncertainty perception(SSBUP)was proposed to solve the problem of single channel Speech separation with background noise.Because the noise is often non-stationary and has strong diversity,the difference of noise distribution between the training set and the test set brings epistemic uncertainty to the speech separation model,and the separation performance of the model is affected.Aiming at this problem,this paper based on the time-domain audio network separation,transform domain characteristics of the mean shift is used to measure forecasting uncertainty,when the uncertainty of the test signals exceeds the threshold,the proposed method can adaptively update the parameters of the coding network in the form of a closed solution without backpropagation,which can reduce the mean deviation of the training and test noises in the feature space and reduce the epistemic uncertainty of the model.Experiments on seventeen thousand mixed data sets on Librispeech,Noisex and Nonspeech public datasets show that the proposed method can improve the separation quality of the models and the intellibility of speech.(2)A dual link speech separation method based on elastic weight consolidation strategy is proposed.In this method,SSBUP method is optimized and improved from two aspects of network structure and parameter updating method.In terms of parameter updating,SSBUP method treats all parameters that need to be adjusted equally,but the importance of coding features varies.To solve this problem,we introduce the elastic weight consolidation strategy,construct the Fisher information weighted Frobenius norm regularization constraint,and update the parameters of the noise coding subnet in time while keeping the important parameters unchanged as far as possible,so as to learn the adaptive transformation domain which is more conducive to the target signal to achieve speech separation.In the aspect of network structure,we adopt the dual link architecture to learn the codec subnet and separate subnet of noise and speech source components respectively,so as to obtain more compactrank expression of speech and noise signals.In this paper,experiments are carried out on Librispeech,Noisex and Nonspeech three public data sets.The experiment includes 16 different kinds of noise mixed data such as animal,bell and machine,and the scale of experimental data reaches 26,000.The experimental results show that the dual link network structure and parameter updating method based on elastic weight consolidation strategy can improve the separation performance of the model.