Lie detection from voice and heart rate

A number of recent studies have shown that much of the information obtained from voice data has a variety of applications in research, including emotion analysis, estimation of gender and age, and lie detection. The aim of this study is to detect a lie by using voice and heart rate data and to compare accuracies of different machine-learning algorithms. Although there are many studies on lie detection in the literature, our study focuses on lie detection from not only voice but also heart rate data. 15 women and 15 men speakers were selected for this study. Each speaker was asked 10 questions and a total of 300 sound files were recorded using an external professional microphone. A smart wristband was used to collect heart rate data from each speaker. Although our main focus was voice and heart rate, a personality test was also given to the participants. Twelve features from the voice data were extracted by using MFCC. Minimum, maximum and average heart rate were calculated for each sound file. Lie detection results calculated from support vector machine, artificial neural network, logistic regression and decision tree methods were compared in the current study. The best accuracy (66,67%) was achieved with Logistic Regression which had 19 features among other used methods. When accuracy rates of all methods were taken into consideration, it was concluded that the better accuracy rate was achieved after the addition of new features which are heart rate, gender and personality type to sound data for lie detection. The main finding of this study was that only the voice data and the pulse data failed to yield any reliable results in detecting the lies. However, it was found that every new added attribute could not improve the results. As a result of this study, we have learned the importance of the selection of attributes and of using the attributes together.