Research And Implementation Of MELP Ultra-low Bit Rate Vocoder

With the fast development of mobile communication technology and the growing amount of users,the bandwidth resource in wireless communication system is becoming more and more valuable.The utilization rate of the frequency-band resource can be improved by reducing the coding rate of the speech.So low rate speech codec with high quality is widely used in various wireless communication systems.Based on linear prediction model,Mixed Excitation Linear Prediction(MELP)vocoder uses five mechanisms such as mixed excitation to improve the quality of synthetic speech,thus high quality synthetic speech can be reconstructed at a rate lower than 2.4kbps.It can be used in satellite communications,military communications,underwater communications and other areas whose bandwidth resources are extremely valuable.Therefore,ultra-low rate vocoder based on MELP is a key research direction in the field of speech codec.As a high quality data compression algorithm,vector quantization plays an important role in ultra-low bit rate speech codec domain.This thesis researches on vector quantization algorithm and proposes GMM-based Predictive Switched Split Vector Quantization(GMM-PSSVQ)algorithm by putting split vector quantization into Predictive Split Vector Quantization(PSVQ).This thesis uses the GMM-PSSVQ algorithm to quantize the Line Spectrum Frequency(LSF)parameter in 2.4kbps MELP vocoder,and compares it with both Multi-Stage Vector Quantization(MSVQ)and PSVQ.Experimental results show that the vocoder which quantizes LSF parameters with GMM-PSSVQ algorithm has a minimum average spectrum distortion and a maximum value for Perceptual Evaluation of Speech Quality(PESQ).It is proved that GMM-PSSVQ algorithm can effectively reduce the quantization distortion of LSF parameters and improve the quality of synthesized speech.Based on the study of the standard MELP algorithm,this article uses the methods of multi-frame joint quantization and linear interpolation,then proposes an ultra-low bit rate vocoder algorithm based on MELP at the encoding rate for 600 bps.In the vocoder,every 20 ms is seen as a frame,and 5 consecutive sub-frames form a super frame.The super frame is divided into 16 modes according to the results of the unvoiced/voiced decision,and the speech feature parameters for each mode are joint quantized with 60 bits.In the allocation scheme for the bit of speech feature parameters,the encoder quantizes only 2 or 3 sub-frames’ LSF parameters with GMM-PSSVQ algorithm in a super frame.After decoding the LSF parameters of these frames in the decoder,the correlation between adjacent frames and the Lagrange interpolation method are used to calculate the unquantized frames’ LSF parameters.The PESQ and Diagnostic Rhymer Test(DRT)are used to test the algorithm.Experimental results show that the speech synthesized by the vocoder proposed in this thesis has high clarity and intelligibility.