在CELP中的降頻取樣碼簿搜尋與修改碼簿

Abstract

CELP(碼簿激發線性預測)是一種利用ABS(合成分析法)而可將語音壓縮到低位元率的演算法。但是，其高運算量在實際應用上總是個問題，而其高運算量主要是在合成分析法所構成的封閉迴圈中。在此論文，我們提出降頻取樣碼簿搜尋與修改碼簿來減少在FS1016裡減少封閉迴路的搜尋的運算量。然而，這個方法會導致音質的下降。所以，我們又提出再搜尋及分類法來改善音質變差這個缺點。經過模擬，我們所提出的方法可以大大降低運算量，同時，所呈現的音質非常接近FS1016。

CELP is a low bit rate speech compression algorithm that applies the ABS (analysis-by-synthesis) method. However, the high computation is always an issue in applications, and the main computation lies in the close-loop in the ABS. In this thesis, we propose the downsample method and modified codebook to cut down the computation in the close-loop search based on FS1016 standard. However, this method would suffer a little degradation in speech quality. Thus, the re-search and classifying method has adopted as a remedy. Finally, the proposed method can reduce the computational load drastically, but the speech quality is nearly the same with standard FS1016.