Analysis And Research On A New Method Of Wavelet Packet Speech Noise Reduction

In recent years,with the rapid development of science and technology,the field of artificial intelligence has become the most popular field of research.Interactive voice system as the most important research content of people's life,the world's major Pinnacle Technology Corp have invested huge funds for research and development.In the process of research and development,the first thing to solve is the problem of voice noise reduction.Wavelet packet transform as the inheritance and development of wavelet transform,multi-scale resolution,selection based flexible,time domain and frequency domain localization properties to the relevance of features to the signal through the multilayer decomposition,the noise in different layers can have good discrimination,related research has been widely used in speech signal denoising.This dissertation investigates the application of wavelet packet transform in the field of speech noise reduction.In view of the problems found in the research,the new preprocessing model and threshold function model are designed from four aspects,and the validity and practicability of the model are tested by simulation.The main contents are as follows:Analysis of the distribution characteristics of wavelet packet decomposition coefficients in order of frequency,frequency according to the sequence of decomposition coefficients are reordered,and according to the signal characteristics of the packet,to solve the traditional noise reduction process in the high and low frequency signal processing problems;wavelet packet decomposition coefficient of energy spectrum based design,a coefficient the pretreatment model,when faced with the original pure voice signal by the noise signal of serious pollution,can be eliminated in advance a portion of the noise signal,keep the original signal to the maximum extent;according to the characteristics of each threshold criterion and application environment,the design of a hybrid threshold processing method,in order of frequency as the basis,according to the characteristics of the signal.The appropriate group of wavelet packet coefficients,different groups in the selection of the best threshold criteria,not only give full play to the the advantages of a threshold rule itself,but also as far as possible to avoid the threshold processing;facing the hard threshold function,there is the step of soft threshold function has constant error problem,design a new threshold function model based on Hermite interpolation function,the model is based on the hard threshold function,not only to ensure that the hard the threshold function advantages,but also overcome the problem of the step,the function nodes in the continuous differentiable,based on the related parameters,can also realize the conversion between the soft and hard threshold function.Selection of pure speech sequence and long short speech sequence,through simulation experiments,the results show that in general a lot of noise,when the noise signal into stationary signals,the new de-noising method can give full play to its characteristics,can not only remove the noise effectively,reduce the pollution of noise signal of the original clean speech signal,but also as far as possible may retain the original signal,reduces signal distortion and signal enhancement of the signal-to-noise ratio and improve the recognition rate of pure speech signal,achieves the noise control effect;when the noise signal is a non-stationary noise signal,a new denoising method has obvious advantages,can reduce the noise of the original speech signal effectively,maximize the recovery of the original pure signal waveform.The experiment fully validates the effectiveness,superiority and practicability of the new denoising method in practical application.