臺灣博碩士論文加值系統

本篇論文主要是探討337個一聲的國字語音之特定語者的單音辨識，研究方向將從337取50個國字語音的辨識出發，再逐步增加字彙量，目的是在辨識率8成的前提下擴充詞彙的數目到中字彙。
論文中所使用到的辨識方法為「K-means之混合式辨識法」及「K-means and 主成分分析法之混合式辨識法」，比較訓練語音之1.向量維度、2.K-means分群數，3.特徵向量個數等因子對待測語音在辨識上的影響及結果。
本實驗結果是使用倒頻譜參數與差倒頻譜參數合併後的參數當特徵參數，進行特定語者的單音辨識，實驗數據在50個國字語音的前提下，最高可得出95.33％的辨識率，字彙量增加到130字時，辨識率還可以維持在84.47％的水準，最後再針對一些可以改進辨識率的方法提供建議。

The thesis is to investigate the speech recognition of 337 isolated mandarin words for the specific speaker. The study will start from 50 isolated mandarin words, and gradually increase to 337 words. We hope that the recognition rate would be at least 80% when the number of mandarin words increases.
The recognition methods we’re using in the thesis are “Hybrid method with K-means” and “Hybrid method with K-means and PCA”. Three factors are considered such as dimension of speech extraction, the number of cluster and the number of eigenvector.
The result of the experiment is operated by speech extraction that is the combination of parameters between cepstrum coefficient and δ- cepstrum. The recognition rate of experiment may highly result in 95.71% under 50 isolated mandarin words. However, the rate of recognition may attain at least 84.46% up when the words increase to 130. Finally some suggestions are given to improve the recognition rate for the future work

論文摘要......................................i
Abstract.......................................ii
目錄.............................................iii
圖目錄..........................................vi
表目錄......................................... viii
第一章 緒論..................................1
1.1 研究動機及目的.......................1
1.2 語音專有名詞簡介....................1
1.2.1 子音和母音...........................1
1.2.2 音頻.....................................2
1.2.3 遮蔽效應(masking effect)........2
1.2.4 訓練語音與待測語音...............2
1.3 國內外相關研究........................3
1.4 語音辨識介紹...........................4
1.5 語音辨識的研究限制.................4
1.6 語音辨識方法概述....................5
1.6.1 語音前處理及特徵向量的求取.5
1.6.2 音框數壓縮與擴展.................8
1.6.3 訓練語音模型........................8
1.6.4 辨識比對..............................9
1.7 論文架構................................10

第二章 語音訊號前處理與特徵向量求取.12
2.1 前言........................................12
2.2 語音訊號前處理........................12
2.2.1 數位取樣...............................12
2.2.2 標準化..................................13
2.2.3 端點偵測...............................14
2.2.4 切割子音與母音.....................15
2.2.5 切割音框...............................16
2.2.6 預強調..................................16
2.2.7視窗化...................................17
2.3 特徵向量的求取........................18
2.3.1 用線性預估係數導出倒頻譜參數.18
2.3.2 差倒頻譜參數(δ- cepstrum)......21

第三章 語音模型的建立與辨識方法...23
3.1 前言.........................................23
3.2 線性壓縮擴張法........................23
3.3 建構模型的方法.........................25
3.3.1 K-means方法.........................25
3.3.2 主成份分析(PCA)....................26
3.3.3 混合式法建立模型...................30
3.4 辨識比對...................................35

第四章　實驗操作流程與實驗結果....40
4.1 操作介面..................................40
4.2 實驗流程..................................40
4.2.1 語音的來源............................40
4.2.2 影響辨識率的可能因子............40
4.2.3 辨識結果................................41
4.2.4 增加字彙的辨識結果...............45
第五章　結論與建議.......................47
附錄..............................................50

[1] 李宗寶，張國清。“用K-means之動態時間軸校正法於國語數字之語音辨識”。 碩士論文，國立中興大學應用數學研究所，台中，2005。
[2] 李宗寶，吳宗憲。“探討K-means之共同向量法應用於國語數字辨識”。碩士論文，國立中興大學應用數學研究所，台中，2005。
[3] 李宗寶，王奕凱。“利用混合式之辨識法於國語數字”。碩士論文，國立中興大學應用數學研究所，台中，2006。
[4] 王小川。“語音訊號處理”。台北市：全華，2004。
[5] Gulmezoglu, M. B., Dzhafarov, Vakif and Barkana, Atalay “ A novel approach to isolated word recognition”, IEEE Trans. On Speech and Audio Processing, vol. 7. No. 6, 1999.
[6] Bilginer Gulmezoglu,M., Dzhafarov, Vakif and Barkana, Atalay ,“The common vector approach and its relation to principal component analysis”, IEEE Trans. On Speech and Audio Processing, vol. 9, No. 6.
[7] Keskin,M.,Gulmezoglu,M.B., Parlaktuna,O. and Barkana,A. ,“Isolated word recognition by extracting personal differences”, in Proc. 6 th Int.Conf. Signal Processing Applications and Technology, Boston, MA, pp.1989-1992, 1996.
[8] Yucel, S., “Application of Gram-Schmidt orthogonalization method to speech recognition for different noise levels”, graduation project, Elect. Electron. Eng. Dept., Osmangazi Univ., Eskisehir, Turkey, 1996.
[9] Angm, H.,“Common vector obtained from linearly independent speech vectors by using LPC parameters”, graduation project, Elect. Electron. Eng. Dept., Osmangazi Univ., Eskisehir, Turkey, 1995.
[10] Bogert, B. P., Healy, W. J. R., and Tukey, J. W.,“ The frequency analysis of time series for echoes : cepstrum , pseudo-autocovariance , cross-cepstrumand saphe cracking ”,in Proc. Symp. Time Series Analysis.New York: Wiley, 1963, pp.209-243.
[11] Bing, X., Yihe, S., Research on ASIC for multi-speaker isolated word reconition, ASIC, 2nd International Conference, 21-24, 135-137,1996.
[12]Tiemey, J., “A study of LPC analysis of speech inadditive noise ”,IEEE Trans. Acoust. Speech Signal Process. ASSP-28 (4) pp. 389-397, 1980.
[13] Atal, B. S., Hanauer, S. L. “ Speech analysis and synthesis by linear prediction of the speech wave ”, J. Acoust. Soc. Am. 50 pp. 637-655, 1971.