Research On Speech Enhancement Of Implantable Middle Ear Hearing Device Based On Deep Neural Network

Implantable middle ear hearing device(IMEHD)has become a hot spot in the research as a new type of hearing aid devices.The IMEHD convert the sound signal into an electrical signal through the microphone firstly,then using digital signal processor to make some process,finally,the actuator converts the received electrical signal proportionally into a vibration signal and transmits it to the inner ear to evoke the sense of hearing.Among them,the signal processing algorithm is especially critical,and the performance of the speech enhancement at the front end of the signal processing directly affects the property of other subsequent algorithms.Considering the traditional speech enhancement algorithms always generate more residual noise,and they are extremely difficult to process the noisy speech with low signal-noise ratio(SNR)and non-stationary noise,the generalization of the noise to be processed is also poor.This thesis was carried out to study the speech enhancement methods of IMEHD,and proposed the application of deep neural network(DNN)to IMEHD's speech enhancement algorithm,its feasibility and performance was also tested.More details are as follows:1)Determining the type of DNN and speech signal feature used in the proposed IMEHD's speech enhancement method.The fully connected DNN was selected as the IMEHD's speech enhancement model,in order to solve the problems of gradient dispersion,over-fitting and local optimization which are often occurred in training DNN using back propagation method,the pre-train method based on restricted boltzmann machine was proposed and was described in detail.Based on the description of the speech generation mechanism and the human ear perception characteristics,the combined features that could better reflect the speech characteristics were determined.2)As part of the proposed IMEHD's speech enhancement method,voice activity detection and noise classification algorithms were designed and implemented.To solve the problem of low recognition accuracy under low SNR conditions,a voice activity detection method based on DNN was proposed.At the same time,the DNN was also used in noise classification.Finally,using the corpus and noise database tested the performance of the proposed algorithms.The result shows that the proposed algorithms achieve better effects and are superior to the traditional methods.3)As part of the proposed IMEHD's speech enhancement method,speech denoise algorithm was designed and implemented.In order to make up for the shortcomings of traditional speech denoise methods,such as easy to leave residual noises,extremely difficult to deal with non-stationary noises,etc.DNN was used in speech denoise algorithm.Using the effective training objectives and comprehensive quality evaluation index of speech,the performance of the proposed algorithm was tested in the selected corpus and noise database.The result shows that the speech denoising using DNN can achieve better performance which does not depend on the choice of corpus,so its applicability is good.Moreover,the proposed speech denoise method is better than the same algorithm without noise classification..4)Experiments were carried out on the proposed IMEHD's speech enhancement method.According the way IMEHD works,a experimental bench was simpledesigned and built,which used for the feasibility and performance testing of the proposed IMEHD's speech enhancement algorithm.It included the dynamic characteristics test of the piezoelectric stack,which selected as the IMEHD's actuator,and the experimental research on the whole set of speech enhancement algorithm.The result shows that the output of selected piezoelectric stack has good stability and fidelity,it can work as the IMEHD's actuator.And the whole set of speech enhancement algorithm can be used for speech signal processing in IMEHD,and the performance is excellent.