Research On Ultra Low Bit Rate Speech Coding

In today’s digital communication networks, speech coding is one ofthe most basic fields. In order to meet the various need in the fields ofdigital communication, the2.4kbps and even lower bit-rate speech codingalgorithm becomes one of the most important research topics in the area ofspeech coding. Especially in some band-restricted situations, ultra low bitrate speech coding algorithms (such as0.6-0.3kbps) are in need.Speech can be decomposed into LP parameters and the residual signalthrough linear prediction coding. The LP parameter is usually converted toLSF parameter before it is quantized. The LSF parameter coding streamoccupies most of the speech coding stream, so its efficient quantizationmethod is an important research topic in the ultra low bit rate vocoder.Firstly, a detailed study is made on the segment vocoder in the topic,and a kind of real-time speech segmentation algorithm is proposed whichis more efficient and reliable. And the algorithm is an important tool of thenew vector quantization algorithm and ultra low bit rate vocoder presentedbelow.Secondly, the new speech segmentation algorithm is combined withthe vector quantization algorithm and a new algorithm called segmentalpredictive multistage vector quantization is designed, in which vectorpredictor is classified into two kinds: inter-segment predictor andintra-segment predictor, which makes a better use of inter-framecorrelation of LSF parameters. Compared to the general predictivemultistage vector quantization algorithm, the new algorithm has anadvantage on the quantization accuracy.Thirdly, a joint optimization training algorithm of code-book isdesigned, which jointly updates the quantizer, the inter-segment predictor and the intra-segment predictor by iteration. The optimized code-book hasa better performance on the quantization of LSF parameters.Finally, the author combines the new algorithms presented above andthe sinusoidal excitation parameter model together, to code and decodeLP(LSF) parameters and the residual signal of input speech, and thenreconstructs the original speech, during the process of which, an ultra lowbit rate vocoder whose average coding rate is about0.4kbps is designed.And the reconstructed speech can be understood.