A study on fast search algorithms using residual signals prediction for G.723.1 speech codec

• Main Authors: Y.C.CHEN, 陳穎澈
• Other Authors: R.S.Lin
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/28081381829114229615

碩士 === 南台科技大學 === 資訊工程系 === 96 === Voice over IP(VoIP), which integrates Internet and speech compression technology primarily, is one of the most popular issues in recent years. A significant part of the VoIP implementation is the low-bit rate speech codec, whose compression rate and computational complexity may sometimes decide the smoothness of an IP call ,especially when the bandwidth of the packet networks is limited. On the limited bandwidth, if the speech compression technology is better, more people can use the band at the same time. Many speech compression technology and researches have been developed and recommended by International Telecommunication Union (ITU-T) for different applications. The recommended G.723.1 is applicable standard codec with low bit rate and good speech quality. In this codec; for quasi-periodic signal, a closed loop fifth orders pitch predictor is computed. And the non-periodic component of the excitation is approximated. For 6.3 kbps, Multi-pulse Maximum Likelihood Quantization (MP-MLQ) excitation is used. Although the codec structure can achieve toll speech quality and low bit rate coding, they required much computational complexity. In this paper, we delivered two major research topics for reduction computation complexity of the codebook search for the G.723.1 codec. First, we propose a fast search algorithm to achieve high efficiency G.723.1 coding. We adopt to predict the search range of adaptive codebook-gain in the G.723.1 standard codec by minimizing the mean square error between the three-tap excitation signal and residual signal. Restrict the search range to reduce the computation required in the five-tap adaptive gain codebook search algorithm. Secondly, we adopt to predict the predetermined excitation codevector and predetermined candidate pulse positions for every subframe before the MP-MLQ search procedure, so the process can reduce computation complexity. Finally, simulation results show that the proposed methods can reduce the amount of computation, and reconstructed speech signal can still maintaining a certain level of speech quality with perceptually negligible degradation.