The Research of VQ-Based Fast Search Algorithm

• Main Authors: Shun-Chieh Chang, 張舜傑
• Other Authors: Shaw-Hwa Hwang
• Format: Others
• Language: en_US
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/k7gsrf

博士 === 國立臺北科技大學 === 電機工程系博士班 === 100 === This dissertation proposes a fast search algorithm for vector quantization (VQ) based on a fast locating method, and uses learning and trade-off analysis to implement this algorithm. The proposed algorithm is a binary search space VQ (BSS-VQ) that determines a search subspace by binary search in each dimension, and the full search VQ (FSVQ) or partial distance elimination (PDE) is subsequently used to obtain the best-matched codeword. In trade-off analysis, a slight loss occurred in quantization quality; however, a substantial computational saving was achieved. In learning analysis, the BSS was built by the learning process, which uses full search VQ (FSVQ) as an inferred function. The BSS-VQ algorithm is applied to the line spectral pairs (LSP) encoder of the G.729 standard, which is a two-stage VQ encoder with a codebook size of 128 and two small codebook sizes of 32. In addition, a moving average (MA) filters the LSP parameter beforehand, and the high correlation characteristics are degraded between consecutive speech frames. These factors present a challenge for developing fast and efficient search algorithms for VQ. In the experiment, the computational savings of DITIE, TSVQ, and BSS-VQ are 61.72%, 88.63%, and 92.36%, respectively, and the quantization accuracy of DITIE, TSVQ, and BSS-VQ are 100%, 26.07%, and 99.22%, respectively, which confirmed the excellent performance of the BSS-VQ algorithm. Moreover, unlike the TIE method, the BSS-VQ algorithm does not depend on the high correlation characteristics of signals to reduce the amount of computation; thus, it is suitable for the LSP encoder of the G.729 standard.