摘 要 本論文主要目的在研究語音基頻週期偵測方法。由於基頻週期為存在於有聲語音訊號中一個相當重要的特徵,在許多語音之處理應用,都需要使用正確的基頻週期特徵。在過去數十年的研究中,提出了很多種基頻週期偵測方法,大多都是使用固定框架長度的架構。之後則有人提出基於變動框架長度的偵測方法,這種方法可以克服固定框架長度之一些缺點,並得到更好的準確度。本論文係針對這種偵測方法,做深入的探討,並提出改善之方法。 我們模擬Xiaohu Qian和Ramdas Kumaresan所提出以變動框架長度為基礎的基頻週期偵測方法,然後針對各種可能之誤判,提出改善的方法,同時我們也探討加入中心截波(Center Clip)的方式所產生之效果。我們使用MAT-2000資料庫來做驗證,比較所提出改善的方法與原本的方法所產生之差異,我們從實驗數據觀察,所提出之方法可以得到更好之偵測結果。
ABSTRACT The objective of this thesis is to study the pitch detection methods of speech signals. Pitch is a very important feature in voiced speech and it is widely used in speech applications. Most of the pitch detection methods are based on fixed-length frames. In 1996, Xiaohu Qian and Ramdas Kumaresan proposed a detection method which was based on variable-length frames. According to the author’s research, the proposed method can overcome some weaknesses of other methods based on fixed-length frames, and can get more accurate results. In this thesis, we study the Xiaohu Qian and Ramdas Kumaresan method and try to improve their algorithm. We also study the effects of applying single clipping techniques in the algorithm. In our simulations, we used speech samples from MAT-2000 and compare the results of different algorithms. From the simulation results, we see that the proposed method can further improve the performance.