Research On Single Channel Speech Enhancement Algorithm Based On Supervised Learning

Speech enhancement is an important technology involving human-computer interaction and speech recognition.In practical application environment,speech is susceptible to noise and reverberation,which results in the decline of speech recognition performance.With the continuous development of the supervised learning model training algorithm,the speech enhancement algorithm based on supervised learning shows a greater advantages than the traditional speech enhancement method.Based on supervised learning,this paper studies the speech denoising and dereverberation algorithms based on supervised learning in noise and reverberation environments,and implements them in the AGV speech recognition control system,which has high theoretical significance and engineering application value.Firstly,the research status of speech denoising and dereverberation technology at home and abroad is described.The model training method of supervised learning and the theoretical basis of speech denoising and dereverberation are introduced.The overall scheme of speech recognition based on supervised learning in noise and reverberation environment is designed.The shortcomings of common speech denoising and speech dereverberation algorithms are emphatically analyzed.It is clear that the key content of this paper is speech denoising and speech dereverberation based on supervised learning.Secondly,in order to solve the problem of noise residue in the low SNR unsteady noise environment of NMF speech denoising algorithm in supervised learning,a single channel speech denoising algorithm based on PM-RNMF was proposed.The algorithm applies psychoacoustic masking characteristics to NMF speech denoising algorithm.Different masking thresholds are applied to different frequency bits.The residual noise energy and speech distortion energy are constrained by the threshold.At the same time,the NMF algorithm is reconstructed by combining the perceptual gain correction with SPP.The experimental results show that the PESQ and SDR values of enhanced speech obtained by PM-RNMF algorithm are higher than those of NMF,RNMF and PM-DNN.Then,in order to solve the problem that the complex non-linear relationship between reverberation signal and clean speech can not be well described by the NMF algorithm and the deep neural network algorithm has a large amount of training data under reverberation conditions,a speech reverberation algorithm based on DNN-NMF isproposed.In this algorithm,the activation coefficient matrix of reverberation signal is obtained by pre-processing reverberation signal with NMF algorithm.Then,the activation coefficient matrix of reverberation signal is re-optimized as the input of deep neural network.Finally,the speech signal is reconstructed by using the optimized activation coefficient matrix.The experimental results show that the DNN-NMF algorithm has better dereverberation performance than DNN,SNMF and LSTM-NMF.Finally,the AGV speech recognition system in noise and reverberation environment is constructed to realize the AGV speech control instruction recognition experiments in different noise and reverberation environments respectively.The experimental results show that PM-RNMF denoising algorithm and DNN-NMF dereverberation algorithm have better speech recognition performance in different noise environments and different reverberation environments,respectively.