結合類神經網路與隱性馬可夫模型之語音辨識系統

Abstract

利用最大可能性訓練程序的隱性馬可夫模型是目前處理語音信號最普遍的 方法．然而它忽略了其它競爭者的訊息，使得系統對於字彙彼此間的區別 能力降低．為了克服這個問題，我們引入類神經網路於原系統中．此結合 隱性馬可夫模型與類神經網路的方法可評估每一字彙於目前發音下出現的 機率而非傳統評估此發音於不同字彙下產生的機率，因而提高辨識系統的 區別能力．本論文的理論架構來自監察式學習與加強式學習兩種訓練機構 ．實驗結果亦證明此方法於語音辨識結果有極大的改善． Hidden Markov Models (HMMs) are currently the most popular method for processingspeech signal. Most often, the maximum likelihood estimation procedure is usedto train the HMMs. Although the maximum likelihood approach has yielded manyencouraging results in practice, it suffers from a serious conceptual problem:it does not include any means of mutual discrimination among word models. In response to this fundamental problem with the HMM, an artificialneural network (ANN) called the multilayer perceptron is incorporated into therecognition system. The hybrid HMM-ANN approach estimates the {\it aposteriori} probabilities from the likelihoods so that the descriminative capability of the recognition system is enhanced. Theoretical justification of the proposed approach is provided and two training algorithms, originated from supervised and reinforcement learning paradigms, are devised. Computer experiments demonstrate that the results of the hybrid HMM-ANN approachto speech recognition are enormously encouraging.