國語語音辨識方法之研究

Abstract

在這篇論文中,我們以文獻所提國語語音辨識之方法做一基礎研究,並建立 一套特定語者之辨識系統．在國語四聲辨識方面,我們採用基頻輪廓 (Pitch Contour) 及能量輪廓為語音特徵,用類神經網路中的多層感知器 (Multi-Layer Perceptron) 來做辨認,得到96.45%的辨識率．在國語408 音辨識方面,我們採用前15個倒頻譜係數(Cepstrum)為語音特徵,用隱態馬 可夫模型(Hidden Markov Model) 來做辨認,在訓練資料量不足的情況下, 仍有76.04%的辨識率． In this paper, we study the Mandarin speech recognition method and establish a speaker dependent recognition system for for isolated words. There are mainly two parts in the system : four tones and 408 syllables recognition. In the four tones recognition , we select the pitch contour and energy contour as the features of speech and apply the Multi-Layer Perceptron for recognition. The results show that the recognition rate of 96.45% can be achieved. In the 408 syllables recognition,we select the first 15 cepstrum coefficients as the features of speech and apply the Hidden Markov Model for recognition. In spite of deficiency of training data, the recognition rate of 76.04% is still obtainable.