利用隱性馬可夫模型與類神經網路於中文單音辨認之研究

Abstract

語音系統利用隱性馬可夫模型辨識是目前處理方法中最受歡迎之一。由於 此法引用最大 可能訓練程序，而忽略了其他競爭者的訊息，使得系 統對於字彙彼此間的區別能力降低。 為了解決這個問題，我們引用類 神經網路，將由隱性馬可夫模型所產生的事前機率值轉 成事後機率 值，增強字彙分辨能力，因而提高辨識率。論文中用監督式學習來訓練網 路 架構，讓每個字模型都具有排它性。實驗結果證明這種方法於語 音辨識（第一候選人） 上有極大的改善。 In this thesis, an automatic speech recognition (ASR) system for isolated Mandarin word is proposed based on hidden Markov models (HMMs) and artificial neural networks (ANNs). The basic processing modules include signal preprocessing, feature extraction, vector quantization, modeling, and classification. In this framework, the HMM serving as the basis is used to model words and recognize utterance. Thus far, the HMM adopting maximum likelihood approach has yielded many encouraging results in practice, but it suffers from poor discrimination among word models. To solve this fundamental problem, we incorporate ANNs into the recognition system. The proposed hybrid HMM-ANN approach uses the a posteriori probabilities instead of the likelihoods so as to preserve the mutual information and offer a mechanism to quantize the discrimination among words. To illustrate the effectiveness of the proposed hybrid HMM-ANN approach, isolated Mandarin word recognition is taken as the examples. All the 408 isolated Mandarin syllables are recognized in our experiment. It is found that the top-one recognition accuracy rate is above 94% for the training set and is above 84% for the testing set. Computer experiments justify that the proposed HMM-ANN approach outperforms the conventional HMM-based approach.