Research On MDCT Non-linear Mapping Model Based Audio Bandwidth Extension Coding

The existing audio bandwidth expansion is divided into a source filter model and a frequency domain model according to the different audio types.This article mainly studies the frequency domain bandwidth extension method for music-like signals.This method uses the principle that there is a correlation between high-frequency signals and low-frequency signals,and the human ear is more sensitive to low-frequency signals.When encoding,first divided the signals into two parts: high frequency and low frequency.and then the low-frequency part is finely encoded,while the high-frequency part only obtains a small amount of additional information to reconstruct the high-frequency signal during decoding;When decoding,first map the low-frequency frequency domain signal to the corresponding high-frequency position,and then use the additional parameters to adjust the high frequency frequency domain signal,and finally get the reconstructed signal.Obviously,this mapping method is based on a linear or quasi-linear relationship between high and low frequency signals,but it is known that audio signals have time-varying characteristics.Therefore,the reconstruction sound qualiy is batter when the correlation between the high and low-frequency signals is strong,and when the correlation become weaker,the reconstruction sound quality will be significantil reduced.In order to solve this problem,wehening the correlation between high and low-frequencies can also maintain a high reconstruction sound quality.This paper establish a bandwidth extension framework based on MDCT nonlinear mapping.The main work is as follows:(1)In the case that the traditional frequency domain rate model uses QMF transformation to cause high coding complexity,a model based on MDCT linear mapping is proposed to improve coding stability and reduce coding complexity.At the same time,before performing time-frequency transformation,the input signal is firstly analyzed by LP to generate an excitation signal,thereby reducing the coding rate.(2)By analyzing the factors affecting the audio quality of the reconstructed high-frequency signals,a quantitative calculation method based on the Euclidean distance and a reconstructed sound quality measurement method based on logarithmic spectral distortion are proposed.The two methods are modeled to verify their accuracy and law.At the same time,it can be used to guide the selection of low-frequency structures for reconstructing high-frequency fine structures.(3)The traditional non-linear mapping function is introduced to analyze the reasons why it is not applicable in the non-blind bandwidth extension technology,and a non-linear mapping model based on tone adjustment is proposed.Compared with the traditional bandwidth extension algorithm,the subjective audio quality is improved by 6.5% and the objective audio quality is increased by 5.7%.(4)Combine the above research,reconstruct the bandwidth extension framework based on MDCT nonlinear mapping.Experimental results show that compared with the SBR method,the subjective audio quality is improved by 2.5%,the objective audio quality is improved by 12.2%,and the bit rate is decreased by 40.1%.Compared with the FBWE method,the audio quality of Guanyin is improved by 13.3%,and the objective audio quality is improved by 23.5%.Compared with the current best e SBR reconstruction method,although the audio quality is slightly lower,the bit rate is reduced by 62.9%.