Time Sequence Features Extraction Algorithm of Lying Speech Based on Sparse CNN and LSTM

Abstract

Time sequence feature extraction algorithm of lying speech based on CNN-LSTM deep network was proposed in this paper. The sparse representation of CNN was realized by introducing $l_{1}$ norm into the objective function of CNN. This sparse optimization algorithm overcame the disadvantage of CNN network that was easy to fall into the local minimum. Firstly, speech preprocessing had been performed, and then, the spectrograms of lying speech were sent into the sparse CNN model. This step was aim to extract the local lying features. Secondly, establishing a time sequence feature extraction model, the local lying features were sent into the LSTM network to extract lying features temporal perspective. Finally, the ${\text{Softmax}}$ testing unit was used to output the lie detection results. Experimental results show that, compared with traditional methods, the model that extracted the fusion features of local features and time sequence features proposed in this paper had a higher detection rate and good scalability. In a word, the Sparse-CNN-LSTM feature extraction model provided a new idea for the research of lying speech detection.