Speech Enhancement Based On Wiener Filter Construction Using Supervision Learning

Nowadays,speech enhancement has achieved a wide range of applications in our daily life.Thus,during the past decades,many different speech enhancement methods have been proposed.Although many supervised speech enhancement methods have been proposed,there are still two problems.First,the speech harmonics contains a lot of useful speech information,so harmonic enhancement is a significantly important task for speech enhancement.However,these methods are difficult to recover speech hormonic structure and remove the residual noise between the harmonic bands.In addition,these supervision learning-based methods usually do not have good generalization.One approach to solve this issue is to increase the number of utterances for noisy speech,noise and clean speech during the training stage so that includes the more noisy environments,but this may be expensive because the parallel database where there is the corresponding noisy speech,noise and clean speech is difficult to acquire.In this work,three strategies are proposed to address the aforementioned problems,which is based on Wiener filter construction and using the supervision learning.Firstly,this work presents a novel codebook-driven speech enhancement method,which applies speech harmonic structure.This algorithm is able to remove more residual noise in harmonic bands because it combines speech harmonic structure and codebook.In this work,we first utilize the speech harmonic structure to estimate prior speech presence probability(SPP)and then use the SPP to appraise the noise autoregressive(AR)spectral shapes for speech enhancement application.In addition,the SPP is also used to modify the Wiener filter.At last,by combining the AR spectral shape codebook of clean speech,we build the modified Wiener filter to acquire the enhanced speech signal.Secondly,this work tries to apply Cepstral Mapping and Deep Neural Networks(DNN)to achieve speech enhancement,which can be more effective to recover speech harmonic structure and gain the higher speech quality.In this work,we apply the DNN to directly predict the Cepstral feature of clean speech and idea Wiener filter given noisy Cepstral input,respectively.Moreover,a fusion framework is proposed to acquire the enhanced speech signal,which combines the Cepstral feature mapping and Wiener filter.Thirdly,we propose a novel framework to conduct speech enhancement,which is based on the long short-term memory networks(LSTMs)and conditional generative adversarial networks(cGANs).This framework includes a generator(G)and a discriminator(D).The G and D are both the LSTMs,so our method is able to be more suitable for speech enhancement task and better recovers speech harmonicity than previous the DNN-based methods.Furthermore,we also apply this network to directly predict the clean speech Cepstral feature and idea Wiener filter given noisy Cepstral input.Additionally,based on the GAN's characteristics,we propose a novel paralleldata-free speech enhancement method based on cycle-consistent generative adversarial network(Cycle GAN),which can reduce the data requirement of the traditional DNNs training.As a result,the more practical noisy speech can be used to increase the generalization of network.