Research On Deception Detection Based On Memory Neural Network

With the rapid development of artificial intelligence and Internet,voice technology has become an indispensable part of our daily life.Among them,deceptive speech detection is a typical application of voice technology in the psychological,judicial and military fields,which is used to detect and judge whether a speaker is lying or not by processing the speaker's voice signal.There are many researches on deceptive speech detection,and most of them are restricted in traditional machine learning algorithms,while the research on deceptive speech detection based on deep learning is relatively scarce.In recent years,the rapid development of deep learning has promoted great progress in various fields,especially the memory neural network has made remarkable achievements in dealing with time sequence problems.This paper mainly studies the deception detection based on memory neural network,and proposes several improved algorithms to improve the recognition rate of the model.The main work of this thesis are listed as follows:1)The research background of deceptive speech detection as well as the research status and difficulty of deceptive speech detection are described.Several public corpus and the production process of STUDENTS corpus are introduced,deceptive speech features and deceptive speech detection algorithms are discussed.2)All the processes of feature before deception detection are analyzed,including the preemphasis,sub-frame and plus windows of speech signals,and the principle of acoustic features such as short-term energy,formant frequency,short-term zero-crossing rate,pitch frequency,Mel Cepstrum coefficients and linear prediction coefficient.3)Several classical machine learning algorithms of deceptive speech detection are introduced detailedly,including Naive Bayes Classifier,k-Nearest Neighbors and Support Vector Machine,the results of experiment show that Support Vector Machine has the best recognition rate among the traditional machine learning algorithms.In addition,the theory of deep learning algorithm is introduced,which provides theoretical basis for the following chapters.4)A deception detection model based on Convolutional memory neural network is proposed,and the differences between statistical features and frame-level features are analyzed.Since statistical features can not represent the temporal information of speech signals well,frame-level features are used in this paper.The model combines the advantages of convolutional network in extracting spatial features and recurrent network in extracting temporal information so that it can extract temporal and spatial features of speech signals better.The model mainly includes Convolutional Bi-directional Long Short-Term Memory(CovBiLSTM)and Bidirectional Long Short-Term Memory(BiLSTM)layers,in which CovBiLSTM is an improved BiLSTM which modifies Hadamard product into convolution operation to learn spatial information of speech.Compared with traditional machine learning algorithm and standard BiLSTM model,the experiment results indicate that deception detection model based on CovBiLSTM can effectively learn the spatial features and temporal information of speech.In addition,the low-level features are extracted by skip connection operation,which improves the recognition rate.5)A deception detection model based on multi-task memory network is proposed,which can improve the generalization ability of the model by using the principle of multi-task learning.The model mainly includes shared hidden layers and sub-task layers.The shared hidden layer is used to extract the shared representation of multiple tasks.The sub-task layer includes gender recognition,deception recognition and speaker recognition(or pseudo-tag recognition).By comparing with the single task model,the experiment results prove that multi-task model effectively improves the generalization performance of the model,thus improving the recognition effect of the model.