驗證碼(Completely Automated Public Turing test to Tell Computers and Human Apart, CAPTCHA)是一種用來區分使用者身份是人類或是電腦程式的驗證系統，通常由一個主要問題與其他干擾組成，較常見的設計有文字、語音與圖像等型態。本研究將進行語音驗證碼的探討，目前此類驗證碼多用於語音認證或適合盲胞使用的非視覺化環境介面。常見的破解方法是透過自動語音辨識系統(Automatic Speech Recognition, ASR )分析訊號內容並進行猜測。據相關研究發現，現階段的語音驗證碼為了提高破解難度，其設計對人類來說普遍都太難，表示尚有改善空間。
雞尾酒會效應(Cocktail Party Effect)是指人類在吵雜的談話環境中，大腦會優先處理主要的聲音訊號並暫時忽略其他不相關聲音的能力，語音驗證碼的設計若能符合此效應，即可提高人類選擇性注意力的優勢。本研究將以此理論為依據，探討現有語音驗證碼組合中是否包含資訊意義干擾以及男女音調(Pitch)差異對錯誤率與喜好程度的影響。
研究結果顯示，干擾中若含有相關資訊意義確實會提升錯誤率，設計時應盡量避免，但最為關鍵的問題卻是男女音調組合的差異性。音調差異較大的組別(如男性播報+群女干擾)錯誤率明顯較低，且主觀喜好評量的分數也較高；而音調差異較小的組別(如男性播報+群男干擾)則出現最高的錯誤率與最低的主觀喜好評價。建議往後語音驗證碼系統設計時可採用音調差異較大的組合(如女播報員+男聲干擾)，既可有效降低程式破解的機率，亦能符合人類聽覺選擇性注意的優勢，提高辨識度。

CAPTCHA is a verification system used for distinguishing whether user identity is human or program through a main question and other interferences. The common designs are patterns like characters, voices and images, etc. This study attempts to explore the voice CAPTCHA, currently, this type of CAPTCHA is mostly used in voice verification or circumstance appropriate for the blind people. The common cracking method is to analyze content and conduct guessing through Automatic Speech Recognition (ASR). According to relevant study, it finds that the voice CAPTCHA at this stage is too difficult for most human beings so as to improve cracking difficulty, which means there’s still room for improvement.
Cocktail Party Effect refers to the ability of the brains of human beings which will process main audio signals preferentially and ignore other irrelevant ones in noisy environment, if the design of voice CAPTCHA is able to accord with this effect, it will improve the advantage of human beings’ selective attention. Based on this theory, this study explores the existence of information meaning interference in current voice CAPTCHA and the influence of pitch difference between male and female on error rate and preference degree.
The study result shows that if there are relevant information meanings in interference, it will truly increase error rate, which should be avoided as possible in design, but the most critical question is the difference in the pitch combination of male and female. There’s significantly lower error rate in the groups which have great pitch difference (for example, male broadcasting and females interference), and the score of subjective preference assessment is higher; Meanwhile, there’s the highest error rate and lowest score of subjective preference assessment in the groups which have small pitch difference (for example, male broadcasting and males interference). It is suggested to adopt the groups which have greater pitch difference (for example, female broadcaster and male interference) when designing voice CAPTCHA in the future, which can both effectively lower the rate of cracking malicious program and accord with the advantage of selective attention in the auditory system of human beings so as to improve visibility. In addition, the broadcast of main message can more adopt the female voice which is rather sensitive for human reaction.

摘要..............I
Abstract.......III
誌謝.............V
目錄.............VI
表目錄...........IX
圖目錄...........I
第一章 緒論...........1
1.1 研究背景..........1
1.2 研究動機與目的.....4
1.3 研究範圍與限制.....5
1.4 研究架構與流程.....6
第二章 文獻探討........9
2.1語音驗證碼.........9
2.1.1破解語音驗證碼....13
2.2 聲音特色..........14
2.2.1男女聲音的差異....16
2.2.2 人類談話聲干擾...17
2.3 噪音干擾..........18
2.3.1 信號噪音比 (signal to noise ratio, SNR)..19
2.3.2 遮蔽效應(masking effect).................20
2.4 雞尾酒會效應(cocktail party effect).........23
第三章 研究方法與設計.............................25
3.1 實驗架構....................................25
3.2 受試者 ..............................26
3.3 實驗設備與環境...............................27
3.4 實驗因子設計.................................32
3.4.1 自變項....................................32
3.4.2 依變項....................................36
3.4.3 控制變項..................................39
3.5 實驗程序與步驟..............................39
3.5.1 實驗程序..................................39
3.5.2 實驗步驟......................................42
3.6 資料處理與分析...................................43
第四章 研究分析與結果.................................45
4.1 受試者基本資料之敘述性統計.........................45
4.2 錯誤率..........................................45
4.2.1 錯誤種類......................................47
4.3 主觀喜好程度評量..................................50
第五章 研究討論......................................52
5.1 錯誤率之探討.....................................52
5.2 錯誤種類之探討...................................54
5.3 主觀喜好程度評量之探討.............................55
第六章 結論與建議....................................56
6.1 結論...........................................56
6.2 未來研究與建議...................................57
參考文獻............................................58
中文部分............................................58
英文部分............................................59
附錄A - 受試者基本資料調查.............................63
附錄B - 受試者須知與實驗同意書..........................64
附錄C - 總體 主觀喜好程度評量問卷.......................65
附錄D - 受試者實驗組別順序.............................66
表目錄
表2- 1 各網站之語音驗證碼設計..........................13
表3- 1 詳細時間規劃..................................41
表3- 2 各組合對依變項之常態檢定........................44
表4- 1 受試者資本資料................................45
表4- 2 各組合與錯誤率之敘述性統計......................46
表4- 3 主訊息與背景干擾對錯誤率之變異數分析..............46
表4- 4 三種錯誤種類之敘述性統計........................49
表4- 5 三種錯誤種類之顯著性比較........................49
表4- 6 各組合與三種錯誤種類之顯著性比較.................50
表4- 7 各組合在三種錯誤種類之錯誤率.....................50
表4- 8 各組合對主觀喜好程度評量之敘述性統計..............51
表4- 9 各組合對主觀喜好程度評量之無母數檢定..............51
圖目錄
圖1- 1 研究流程圖....................................8
圖2- 1 reCAPTCHA系列之語音驗證碼……………………………………........11
圖2- 2 男女音調差異圖.................................16
圖2- 3 固定頻率純音遮蔽另一純音之示例圖 ...........22
圖3- 1 DUOCAPTURE實體圖片……………………………………………...........28
圖3- 2 鐵三角電容式麥克風實體圖片........................28
圖3- 3 Gold Wave 5.56繁體中文版 操作畫面...............29
圖3- 4 MP3Gain音量校正軟體操作畫面......................30
圖3- 5 實驗人員操控畫面...............................31
圖3- 6 實驗環境畫面...................................32
圖3- 7 實驗環境畫面...................................32
圖3- 8 Yahoo語音驗證碼之音訊頻率分佈圖..................36
圖3- 9 實驗說明階段流程圖..............................40
圖3- 10 正式實驗流程圖................................41
圖4- 1 實驗各組合之錯誤率分布圖……………………………………….......…..47
圖5- 1 人類對訊號處理之認知流程圖………………………………………........54

中文部分
[1] 韓寶強，「音的歷程」，中國文聯出版社，2003。
[2] 蕭雅文，「聽力學導論」，五南圖書出版公司，第8-31頁，2008。
[3] 詹智閔，「小波分析在音樂訊號上的應用」，生醫電資所，2012。
[4] 徐茂銘，漫談聽力及聽力障礙，1977。
http://210.60.224.4/ct/content/1977/00110095/0006.htm
[5] Henry Doktorski，「在線音樂理論:音調、音質與音色」，2012。
http://www.dolmetsch.com/musictheory27.htm
[6] 張智星，「音訊處理與辨識」，2006。
http://neural.cs.nthu.edu.tw/jang/books/
audiosignalprocessing/index.asp
[7] 林信鋒、蔡正富，「植基於離散小波轉換之聲音浮水印技術」，
東華大學，2004。
[8] 劉超群，「雙重任務干擾與資訊處理模式」，國立中央大學機械工程學
系，2010。
[9] 許勝雄、彭游、吳水丕，「人因工程」，滄海書局，第118-119頁，2004。

英文部分
Ahn, L., Blum, M., & Hopper, N. J. (2004). Telling humans and computers apart (Automatically) or How lazy cryptographers do AI.Communications of the ACM, 47, 57–60.

Brokx J.P.L and Nooteboom S.G. (1982) Intonation and the perceptual separation of simultaneous voices. Journal of Phonetics 10,23-36.

Barry Arons (2008) . A Review of The Cocktail Party Effect .MIT Media Lab Conversational Computer Systems.

Bigham, J. P., & Cavender, A. C. (2009). Evaluating existing audio CAPTCHAs and an interface optimized for non-visual use.Proceedings of the 27th international conference on Human factors in computing systems, CHI ’09 (1829–1838)

Cherry, E. C. (1953). Some experiments on the recognition of speech, with one and with two ears. Journal of the acoustical society of America,25(5), 975–979.

Chan, T.-Y. . (2003). Using a test-to-speech synthesizer to generate a reverse Turing test. 15th IEEE International Conference on Tools with Artificial Intelligence, 2003. Proceedings (p 226 – 232).

Chellapilla, K., Larson, K., Simard, P. Y., & Czerwinski, M. (2005). Building segmentation based human-friendly human interaction proofs (HIPs). Proceedings of Second International Workshop on Human Interactive Proofs, May 2005, 1–26.

Cainer, K. E., James, C., & Rajan, R. (2008). Learning speech-in-noise discrimination in adult humans. Hearing Research, 238(1–2), 155–164.

Du, Y., Kong, L., Wang, Q., Wu, X., & Li, L. (2011). Auditory frequency-following response: A neurophysiological measure for studying the “cocktail-party problem”. Neuroscience & Biobehavioral Reviews, 35(10), 2046–2057.

Helenius, R. & Hongisto, V., (2004), The effect of acoustical improvement of an open-plan office on workers. Proceedings of Inter-Noise 2004, Paper 674, 21–25 August.

Haichang Gao, Honggang Liu, Dan Yao, Xiyang Liu, & Aickelin, U. (2010). An Audio CAPTCHA to Distinguish Humans from Computers. 2010 hird International Symposium on Electronic Commerce and Security (ISECS)(pp. 265-269).

Jennifer Tam, Jiri Simsa, Sean Hydn, and Luis Von Ahn.(2007). Breaking Audio CAPTCHAs. Computer Science Department Carnegie Mellon University .

Jennifer Tam, Jiri Simsa, David Huggins-Daines, Luis von Ahn, and Manuel Blum.(2008). Improving Audio CAPTCHAs. Computer Science Department,Carnegie Mellon University .

Killion, M.C., Niquette, P.A., Gudmundsen, G.I., (2004). Development of a quick speech-in-noise test for measuring signal-to-noise ratio loss innormal-hearing and hearing-impaired listeners. J. Acoust. Soc. Am.116, 2395–2405.

Kolupaev, A., & Ogijenko, J. (2008). CAPTCHAs: Humans vs. Bots. IEEE Security & Privacy, 6(1), 68-70.

Sauer, G., Hochheiser, H., Feng, J., & Lazar, J. (2008). Towards a universally usable CAPTCHA. Proceedings of the Symposium onAccessible Privacy and Security.

Soupionis, Y., Tountas, G., & Gritzalis, D. (2009). Audio CAPTCHA for SIP-Based VoIP., Emerging Challenges for Security, Privacy and Trust, IFIP Advances in Information and Communication Technology ( 297, p25–38).

Venetjoki, N., Kaarlela-Tuomaala, A., Keskinen, E., & Hongisto, V. (2006). The effect of speech and speech intelligibility on task performance. Ergonomics, 49(11), 1068–1091.

Wickens,C. D.(1984).Engineering psychology and human performance.Upper Saddle River New Jersey 07458.

Yan, J., & El Ahmad, A. S. (2008). Usability of CAPTCHAs or usability issues in CAPTCHA design. Proceedings of the 4th symposium on Usable privacy and security, 44–52