Research On Speech Bandwidth Encoding Based On Nonlinear Mapping Of Sinusoidal Models

In recent years,with the development of speech signal processing technology,people's demand for speech bandwidth has become higher and higher,and narrowband speech can no longer meet the needs of social development.Speech bandwidth expansion is an important part of modern speech codecs.Studying speech bandwidth expansion is of great significance for improving the quality of speech hearing and reducing the coding rate.This article starts with the study of the relationship between the high and low frequency amplitude parameters of the sine model.Using the theoretical basis of the correlation between high and low frequency signals,the correlation between the high and low frequency amplitude parameters is verified through data statistics.Through LSTMRNN deep neural network model modeling,in-depth mining of the actual relationship between the high and low frequency signal amplitude parameters,an improved speech bandwidth expansion coding method based on sinusoidal model nonlinear mapping is proposed.The main research contents of this article are as follows:(1)Summarizes the generation mechanism and basic characteristics of speech signals,introduces the preprocessing process of speech signals and the commonly used feature parameter extraction methods;studies the analysis and synthesis algorithm of the sine model,and conducts analysis and synthesis of the sine model It gives a detailed introduction and analyzes the influence of the number of sinusoidal parameters on the quality of the synthesized speech,so as to pave the way for bandwidth expansion later.Finally,in order to show that the sine model is a good speech parameter model,it is reconstructed and simulated.(2)The relationship between the high and low frequency amplitude parameters of the sine model is studied.Through the principle analysis of the bandwidth expansion of the sine model,it is considered that there should be a correlation between the high and low frequency amplitude parameters,and then the correlation is calculated by calculating the Euclidean distance between the high and low frequency amplitude parameters and the log spectrum distortion of the high frequency reconstruction quality Verified.(3)According to the conclusion of(2),using the LSTM-RNN deep neural network to model the high and low amplitude parameters of the sine model,an improved speech bandwidth expansion algorithm based on the nonlinear mapping of the sine model is proposed.In order to illustrate the advantages and disadvantages of the algorithm in this paper The performance is compared with the classical sine model bandwidth expansion algorithm and the improved source filter model bandwidth expansion algorithm,and their performance is evaluated through subjective and objective tests.The experimental results show that compared with the classical sine model bandwidth extension method,the algorithm in this paper has been improved in terms of reconstruction sound quality and bit rate reduction;compared with the improved source filter model bandwidth extension algorithm,the two are basically subjective and objective in sound quality Quite,but the algorithm code rate in this paper is reduced by 50%.