中文語音聲調辨識演算法

Abstract

中文是一種有聲調性的語言，所以中文語音辨識系統的準確性會受到聲調辨識演算法的直接影響。聲調辨識的問題在建立一個高信賴度的中文語音辨識系統上是值得仔細探討的。 在這篇論文中，我們研究中文語音辨識的聲調問題。我們設計了四個特徵函數來區分中文的聲調，我們也提出了三種主導區分聲調的方法。在我們做的實驗□，我們用大量的測試資料(每個說話的人有1390個音節, 共19人)來驗證這三種區分聲調的方法。測試結果最佳方法是一種混合法，該法在多層決定樹的第一層用歐幾理得距離算法來增進辨識效果，它來測試男聲和女聲各得到89.35%及80.90%的正確率。而在個人方面該法對其中一位男聲可達到94.7%, 92.94%, 89.7%和92.4%的四聲辨識正確率

The Chinese language is a tonal language. The accuracy rate of a Chinese speech recognition system will be affected directly by the accuracy rate of its tone recognition algorithm. The recognition problem of tones deserves a closer look if we want to build a highly reliable Chinese speech recognition system. In this thesis, we studied the tone recognition problem in Mandarin. We have designed four features to distinguish the tones. We have also developed three major ways to classify tones. We have performed experiments to test the classification methods on large test samples (1390 syllables for each speaker with 19 speakers in total). The best classification method is the mixture one where we use the Euclidean distance on a multi-level decision tree. Its recognition rate is 89.35% for males and 80.90% for females of the mixture method. Individually, the recognition rate of the mixture method can reach higher than 90%. For on male the rate is 94.7%, 92.94%, 89.7% 92.4% for tone 1 to tone 4.