向量和激發線性預測語音編碼法之研究與改進

Abstract

在無線語音通訊系統中，穩定､複雜度低之低率語音編碼法是十分重要 的一部份。本論文探討北美數位行動電話標準 IS-54/136 所採用之向量 和激發線性語音編碼法，並考慮在運算量降低及語音品質提升兩方面改善 編碼器之效能。 在運算量降低方面，我們提出數種方式降低佔 90% 以 上運算量之碼簿搜尋及增益量化等程序。這些方式在少量的 SSNR 損失下 ，降低整體編碼運算量各達 6% 至 20%，並可被同時使用於編碼器之改良 上。在語音品質提升方面，我們提出一種樹狀碼簿搜尋法，可提升 SSNR 約 0.3-1 dB。此外，我們亦提出降低樹狀碼簿搜尋法運算量之有效方式 。對於樹狀碼簿搜尋法之參數､運算複雜度及語音品質改善之間的關係， 我們更作了深入的分析及歸納，並以簡單的數學式表示之。 In wireless communication systems, a robust, low complexity and low bitrate speech coding algorithm is essential for voice applications. Vector-Sum excited linear predictive coding (VSELP) is one of the coding algorithms that can meet the requirement. We study the North American IS-54/136 VSELP speech coding algorithm and consider ways of encoder improvement in the sense of reducing complexity or improving speech quality. In terms of complexity reduction, several techniques are presented for longterm predictor search, excitation codebook search and gain quantization, which occupy more than 90% of encoding complexity. These methods respectively reduce roughly 6% to 20% of encoding complexity at moderateSSNR loss and can be applied to the encoder simultaneously. In terms of quality improvement, a tree-structured codebook search is considered. Roughly 0.3-1 dB SSNR improvement can be achieved by tree-search coding. We also present a technique which can significantly reduce tree-coding complexity. Moreover, we analyze simulation results to obtain rule-of-thumbformulas for determination of the various tree parameters under different complexity and performance conditions.