本論文中，採取MFCC係數作為特徵參數，以隱藏式馬可夫模型作為聲學模型，再以維特比搜尋法計算機率值。為了簡化及減少運算量，我們研究了一些把浮點運算轉換成定點運算的作法。此外再把隱藏式馬可夫模型中的變異數參數去除，也可使維特比搜尋的速度更快。聲學模型之訓練，使用兩階段訓練法，首先以等切和向量量化來訓練初始模型，接著使用分段K中心法再對聲學模型作訓練，這些訓練程序都使用定點運算來進行。
在使用手工標記端點語料的實驗，語者無關的辨識可達到3.57% 的辨識錯誤率。若以程式作自動的端點偵測，則可達到5.92%的辨識錯誤率。經過模擬實驗之後，我們也已經將辨識階段的程式搬到Pocket PC上，實際測試顯示，我們的程式可以正常的執行語音命令的辨識處理。

To simplify and decrease the quantity of computations, we have study some methods to replace floating-point operations with fixed-point operations. In addition, by eliminating the parameter of variance from HMM, the speed of Viterbi search is further increased. For acoustic model training, uniform segmentation and vector quantization are used to train initial model parameters. Then, segmental K-means is used to refine the model parameters. These two training steps are all executed with fixed-point operations.
When the endpoints of all utterances are manually marked, the recognition error rate obtained is 3.6% in the speaker-independent experiment. But, when the endpoints are detected automatically, the error rate obtained becomes larger, i.e. 5.9%. After simulation experiments, we have ported the recognition program to Pocket PC. According to practical testing result, our program can normally recognize voice commands on Pocket PC.

[1] Lawrence Rabiner and Biing-Hwang Juang, “Fundamentals of Speech Recognition”, Prentice-Hall International, Inc.1993.
[2] Sadaoki Furui, B.H.Juang, Wu Chou, “Automatic Speech Recognition and Understanding Proceedings”, IEEE Workshop 1997.
[3] Douglas O’Shaughnessy, “Speech Communications Human and Machine Second Edition”, The Institute of Electrical and Electronics Engineers, Inc, New York 2000.
[4] Alan V.Oppenheim and Ronald W. Schafer with John R.Buck, “Discrete-Time Signal Processing Second Edition”, Prentice-Hall International, Inc.1999.
[5] Brian Delaney, Nikil Jayant,Mat Hans, Tajana Simunic, Andrea Acquaviva, “A Low-Power, Fixed-Point, Front-End Feature Extraction for a Distributed Speech Recognition System”, ICASSP 2002.
[6] Christophe L´evy, Georges Linar`es1, Pascal Nocera and Jean-Franc¸ois Bonastre, “Reducing Computational and Memory Cost for Cellular Phone Embedded Speech Recognition System”, IEEE ICASSP ITT-P1.6, May 2004
[7] 陳松琳，以類神經網路為架構之語音辨識系統，中山大學電機工程學系，碩士論文，2002。
[8] 林育祺，利用可重組織單晶片系統實現語音辨識，私立逢甲大學資訊工程學系，碩士論文，2004。
[9] 楊永泰，隱藏式馬可夫模型應用於中文語音辨識之研究，中原大學資訊工程學系，碩士論文，1999。
[10] 陳明，DSP在語音辨識系統之應用與研究，台灣科技大學，機械工程系，碩士論文，2004。
[11] 楊茗惠，用隱藏式馬可夫方法於頻域特徵之國語數字辨識，中興大學應用數學系，碩士論文，2002。
[12] 王閔鴻，不特定語者大詞彙華台雙與辨識引擎之研製及其應用，長庚大學，碩士論文，2003。