中文語音測謊技術開發初探

A Preliminary Research of Chinese Acoustic Lie Detection Technology Exploitation

為因應國內測謊相關規範以及企業對謊言偵測需求，達到犯罪預防與偵查並重目的，非接觸式測謊技術之使用已成為必然趨勢，本研究基於此概念，開發出本土中文語音測謊雛形。受限於語音資料庫因素，本研究從個人化測謊概念進行設計，模擬試驗中，以不同時序收集27位，每位共六段誠實／說謊語音；實測案例，則由司法機關委託測謊案件中選取50件，根據測前晤談程序內容，區分為中性／相關／比對問題之語音。語音經過剪輯，保留受測者聲音後，以語音活性檢測、openSMILE語音情緒特徵分析與選取等步驟處理，後以區別分析判定各組差異性。結果顯示，僅以基本語音特徵作為說謊判別可能存在盲點；模擬案件之交叉驗證正確率可達約85％，實測案件甚至可達92％，檢視語音特徵分類狀況，可發現作為分類基礎之語音特徵大致相同，但每位語者存在細部差異，且系統選取之特徵會隨著錄音品質而有所不同，錄音品質較佳者可以選取較多特徵，但使用特徵數與正確率並未發現正相關。綜合前述結果，本研究認為，本研究使用的個人化語音分析方式，具有發展成為本土新式測謊系統之潛力，惟如何事先收集到當事人足資比對之語料則為下一步開發重點。

The polygraph test can be used for criminal investigation and prevention. Contactless lie detection has become a trend sooner rather than later in our country due to the requirement of polygraph policy and business management for lie detection. The goal of this research is to create a prototype of a native acoustic lie detection technology based on this concept. Due to the scarcity of a Chinese audio database, this study relies on the principle of personalized lie detection. In a simulation test, 27 participant sample data sets were obtained, containing truth and deception acoustic files in varied chronological sequences. In 50 genuine polygraph cases, audio files from the neutral/relevant/ comparison questions during the pretest interview stage were collected to evaluate the real polygraph cases. Only the subjects’ voices are preserved when the audio files are edited. These files are submitted to voice activity detection, openSMILE feature extraction and selection, and discrimination analysis during the data editing process. The results indicate that identifying the truth purely based on specific auditory features is unreliable. The accuracy rate of corss-validation is 85% in simulation cases and 92% in real cases. By examining the classification results, we can see that the acoustic characteristics used for categorization are roughly the same, although each person has subtle variances. The number of features selected varied with recording quality; more features were picked in higher recording quality files, but there is no positive correlation between the number of features selected and the accuracy rate. Based on these results, we believe that the personalized acoustic analysis methodology utilized in the study has the potential to be used for native lie detection. The subsequent research will concentrate on the subject's standard corpus collection for comparison.