基於語音特徵判斷語句內容真實性

Determining the authenticity of speech content via analyzing the voice characteristics

吳乙彤(Yi-Tung Wu)

指導教授 ： 劉奕汶

國立清華大學/電機資訊學院/電機工程學系所/碩士(2018年)

摘要

現今社會中，詐騙事件層出不窮，其中以電話進行詐騙佔較大多數。如果能對詐騙電話的語音內容進行探討，分析一段虛構內容的語音特徵，進一步判別此通電話之目的真實性，就可能得以幫助預防電話詐騙的發生。截至目前為止，詐騙相關的研究愈來愈受到重視，故我們由國立清華大學的學生中招募有意願參與本實驗的受試者，並設計一份問卷及流程，以遊戲的形式搜集受試者說謊與說實話的語音資料，進一步建立謊言辨識語音資料庫並進行相關研究。本研究針對受試者回答的錄音內容以數位訊號處理方法做語音特徵分析，最後搭配決策樹學習的訓練模型，對一段未知的語音藉由特定的語音特徵辨別出真偽。本研究亦根據個人特徵重要度的不同建構個人化模型以及大眾化模型，使得模型能因應不同人重要特徵不同的差異，進而得到效能較好且一般化能力較高的模型，並嘗試藉此將受試者進行行為群聚性分析。除了列出機器學習的成果並比較特徵選取前後辨識率之差異以外，我們亦根據目前研究的結果，提出未來能繼續改善、增進的方向，例如加入更多特徵如笑聲辨識、聲音明亮度等等，亦會找出實現特徵權重分配更好的方法，以提升大眾化模型之效能。

關鍵字

測謊 ； 決策樹 ； 語音特徵

並列摘要

In the society, scams are everywhere, and the most common way to fraud is phone scam. If we can determine the authenticity of phone call contents by analyzing the characteristics of fake speech, it will help preventing phone scams. So far, deception-related research has received more and more attention. In this research, we recruited the students from National Tsing Hua University to become subjects, and collected speech data containing truths and lies in the form of a game. A questionnaire was designed and processed, so that the ground truth can be labeled for the entire database. Then, we analyze the recorded speech data of subjects by using digital signal processing methods. Finally, using decision tree learning technologies, we aim to develop an algorithm to determine the authenticity of speech content automatically. In our work, we also construct the personal model and the general model based on the importance of individual characteristics, so that the model can adapt the differences between important characteristics of individuals, and then obtain the model with better performance and higher generalization ability. Furthermore, we try to analyze that if subjects’ behavior has a tendency to cluster when they are lying. In addition to listing the results of machine learning tests and compare the difference before and after feature selection, we also put forward the future work according to the current results. One possible direction would be to involve more features, like laugh detection, tone, and so on. Also, it might be possible to search for a better way to implement feature weighting and improve the efficacy of the general model.

並列關鍵字

deception ； speech ； decision tree

參考文獻

[1] P. Boersma, “Praat, a system for doing phonetics by computer,” Glot International 5:9/10, pp. 341-345, 2001.

[2] Praatio:Tim Mahrt. PraatIO. https://github.com/timmahrt/praatIO, 2016

[3] 易作霖1920《國音學講義》，商務印書館。

[4] J. P. Burg, “Maximum entropy spectral analysis,” Annual International Meeting, Soc. of Explor. Geophys., Oct., 1967.

[5] H. Gray and D. Y. Wong, “The Burg algorithm for LPC analysis/synthesis,” IEEE Trans. on Acoust., Speech, and Signal Processing, vol. 28, no. 6, pp. 609-615, Dec. 1980.