Research And Implementation Of Low Rate Speech Encoder Based On CS

Compressed Sensing(CS)breaks people's constraints on the traditional Nyquist sampling theorem and can achieve high-quality signal reconstruction under low sampling rate conditions.The low-rate speech coder adopts parametric coding technology,has a low coding rate,and requires a small transmission bandwidth,and can be applied to fields such as satellite mobile communication,underwater communication,and military communication.Low-rate speech coding technology is one of the effective methods to improve the frequency band utilization in wireless communication.Research and implementation of low-rate speech encoder with higher synthesized speech quality have practical application value.In this thesis,CS theory is applied to the efficient quantization of speech feature parameters,which solves the problem of classification and reconstruction of speech signals with different sparsity degrees,and implements a speech encoder with a rate of 1 kbps.The Linear Spectrum Frequency(LSF)is a channel model parameter important for speech,and the number of quantization bits allocated in the encoding process is the largest.In order to achieve transparent quantization of LSF with as few bits as possible,this thesis proposes an Adaptive Reconstruction Algorithm for Compressed Sensing(ARA-CS).The algorithm firstly uses the CS method to observe the LSF parameters of the speech frame.When reconstructing,adaptively selects the adjustment parameters according to the clear/turbidity type of the speech frame,and determines the dimension of the perceptual matrix according to different adjustment parameter values.The linear equation solving or least squares method obtains the first part of the sparse coefficient of the LSF parameter,and the latter part complements 0;finally,the sparse inverse transform is performed to obtain the reconstructed LSF parameter.The algorithm is also applicable to the observation and reconstruction of continuous multi-frame LSF parameters.The performance of the algorithm is evaluated by the average reconstructed signal-to-noise ratio and spectral distortion performance.The test results show that the proposed algorithm outperforms the traditional CS algorithm for LSF reconstruction of voiced frames and multi-frame joint coding.Based on the MELP coding model,a multi-frame joint coding technique and the ARA-CS algorithm proposed in this thesis are used to design a 1 kbps speech codec.The encoding end divides the speech signal into 30 ms long sub-frames,and successive2 sub-frames form a superframe.According to the clear/turbid type of the sub-frame,the super-frame is divided into four structures,and each super-frame structure is quantized by 60 bits.The ARA-CS algorithm is used to observe and reconstruct the LSF parameters of different superframe types by using different adjustment parameters,and obtain low-dimensional observation sequences,and then perform vector quantization coding.The coded required quantized codebook is also trained by the ARA-CS algorithm and the codebook dimension is adjusted according to the adjustment parameters.The designed encoder is tested by PESQ and DRT.The results show that the average MOS value of the encoder is 2.64 and the DRT score is 87.29%.The synthesized speech has better definition and intelligibility.