Lie detection from speech analysis based on K–SVD deep belief network model

Abstract

Considering the task of lie detection relates some nonlinear characteristics, such as psychological acoustics and auditory perception, which are difficult to be extracted and have high computational complexity. So this paper proposes a deep belief network based on the K-singular value decomposition (K-SVD) algorithm. This method combined the multi-dimensional data linear decomposition ability of sparse algorithm and the deep nonlinear network structure of deep belief network. It is aim to extract the significant time dynamic deep lie structure characteristics. Based on these deep characteristics, the lie database of Arizona University at United States was used to test. The experimental results show that, compared with the K-SVD sparse characteristics and basic acoustic characteristics, the deep characteristics proposed in this paper has better recognition rate. Furthermore, this paper provides a new exploration for psychology calculation.