Novel Lie Speech Classification by using Voice Stress

Felipe Marcolla, Rafael de Santiago, Rudimar Dazzi

Abstract

Lie detection is an open problem. Many types of research seek to develop an efficient and reliable method to solve this problem successfully. Among the methods used for this task, the polygraph, voice stress analysis, and pupil dilation analysis can be highlighted. This work aims to implement a neural network to perform the analysis of a person’s voice and to classify his speech as reliable or not. In order to reach the objectives, a recurrent neural network of LSTM architecture was implemented, based on an architecture already applied in other works, and through the variation of parameters, different results were found in the tests. A database with audio recordings was generated to perform the neural network training, from an interview with a randomly selected group. Considering all the neural network base models implemented, the one that showed prominence presented a precision of 72.5% of the data samples. For the type of problem in focus, which is voice stress analysis, the result is statistically significant and denotes that it is possible to find patterns in the voice of people who are under stress.

Download


Paper Citation