Research On Mixed Excitation Linear Prediction Speech Coding Algorithm

With the development of communication technology and the improvement of the speech transmit technology in real time based on Internet, the high request to transmission efficiency and memory capacity of the speech is needed. One of the main ways to solve these problems is the speech coding. Therefore, speech coding research, in particular the low bit-rate speech coding study is of great practical significance.The Mixed Excited Linear Predictive Coding (MELP) is a good method in the study of the existing speech coding. It combines the advantages of the dual incentive, code incentive and multi-band excitation and divideds into a number of sub-bands for the short-term speech . Each sub-band sound will be discriminated by unvoiced and voiced respectively. It uses the mixed periodic sequence and random noise pulse to motivate speech synthesis filter in the end of the synthesis and can get effectively renewable speech in a lower bit rate. The coding method of 2.4 kbps Mixed Excited Linear Prediction Speech Coding has been established as the new United States Federal speech coding standards.Through researching the algorithm theory of MELP speech code, this paper makes more in-depth research for the process and we also find that the area can makes improvements in the pitch and linear prediction coefficient quantified. In the standard MELP algorithm the value of the relevant is small for the irregular cycles of the speech signal and we mistakenly believe that it is unvoiced. So it will introduce noise. Here adopted a change in the algorithm of pitch and made it calculate more accurate. In addition, the storage and computing complexity are very great in the quantification of LSF. Address this issue, we have proposed a three vector quantization method. The rate of MELP can be dropped to around 2.1 kbps and there are still good quality of synthetic voice. In the end of the article, the algorithm is simulated by the MATLAB programming environment. The simulated results show that the output waveform of decoded speech signal and improved speech signal is very similar with the waveform of the original speech signal, but there are some distortions and there are some time-delays to a certain extent in the segment of voiced. Thereby it can be verified that the algorithm is theoretically feasible.