Research On Low-SNR Speech Enhancement In Driving Environment

With the development of artificial intelligence,intelligent drive has gradually become the focus of people's attention.Intelligent in-car electronics,such as portable navigation devices,smart rearview mirrors,also get into our daily life by degrees.In order to reduce accidents and free driver's hands,most of the electronic products in-car need to be controlled by voice commands.However,driving environment is a complex acoustic environment which is filled with various noises.In such environment,the technology of speech recognition at present is difficult to meet our needs.Therefore,research on technology of speech enhancement in situation of driving is significant for future design and development of automotive electronic products.The speech enhancement algorithm for driving environment with low the signal to noise ratio(SNR)is studied in this paper,and focuses on the technologies based on neural networks.The fundamental and the implementation of the method and their improved forms are presented,following is the main work of the paper:1.According to the characteristics of automobile noise and the driver's voice,the optimization scheme of a deep neural network training algorithm based on restricted Boltzmann machine theory was proposed in the paper.In this algorithm,the original and unsupervised training scheme was changed to a supervised training scheme which puts the higher SNR parallel corpus as the objective function.Compared with the original training algorithm,the proposed scheme not only alleviates the problem of over fitting when the training set data are too small,but also simplifies the structure of the network model and reduces the operation time of enhancement.2.In order to utilize the advantages of Back Propagation(BP)neural network and Radical Basis Function(RBF)neural network,a training algorithm which combines these two kinds of network was proposed in this paper.The training algorithm not only makes up for the shortcomings of nonlinear BP network mapping ability but also minimizes the number of hidden layer neuron of RBF network.3.The optimization program of two aspects is taken in this paper to improve the de-noising ability and robustness of the proposed algorithm.In order to improve the performance of the enhanced model,the training set data are classified according to SNR.In order to accurately and fast identify the strength of signals with noise SNR during enhancing,the Voice Activity Detection(VAD)algorithm based on energy complexity statistics is used to detect the starting point of speech,and then analyzes the pure noise segment before the starting point to determine the categories that noisy speech belongs to.In order to improve the robustness of the algorithm,the input data which is different from the global mean value during the enhancement is conducted in the paper to better adapt to the enhancement model.In the reconstruction process of a speech,in order to reduce the music noise produced by the mean shift algorithm,the minimum controlled recursive averaging noise estimation algorithm is adopted in this paper to estimate the speech presence probability,which are weighted summed based on the speech presence probability.The results of experiment show that the proposed algorithm can be applied to the speech enhancement in driving environment,and it has superior improvement on segmented SNR and speech quality compared with the existing algorithms.