Research Of Lie Detection Based On The Speech Sparse Representation

Signal sparse representation is a kind of compression method that using a few characteristic vectors to represent the original signal.It fuses many disciplines,such as cognitive neurology,psychological physiology and computer science.It is a typical efficient signal representation method at present.Speech has the nature characteristic of sparsity,and the acquisition of speech signal is simple and easy.For these reasons,many scholars usually research the speech signal to obtain the essence structure characteristics by using sparse representation method.It is widely used in many research fields,such as compressed sensing,feature recognition,psychological calculation and so on.Lie speech detection is a process of lying psychological state recognition based on speech signal analysis.Lying psychological state is a dynamic process that is gradient and reciprocating.It involves three parts,including psychological emotion,psychological cognition and psychological will.So lie detection is a complex scientific problem.From the aspect of speech stream,lie is generally hidden in the speech clips and can not be easily found.From the aspect of information stream,lie is the local information in the speech information stream.Normally,the emotion must be tension when people lies.As a result,this tension mood leads to some subtle changes in the sound tack structure,for example,the semantic characteristics,prosodic characteristics,resonance peak and the psychoacoustics parameters all can be different from before.Therefore,the phonetic characteristics that express lying state perform sparse distribution.Because of this,here,the sparse representation method is used for researching the problem of lying psychological state calculation,and this is a reasonable choice.At present,the scholars at home or abroad research the lying speech detection mainly focused on the psychological emotion analysis,and there is almost no research involves the psychological cognition and psychological will.So this paper chooses the complicated problem of lie detection as the oriented research object,and the sparse representation method is used to solve some problems of the lying psychological calculation.The main contributions are as follows:1.A lying speech feature extraction method based on sparse representation is put forward.Because lying speech contains more abundant characteristic information than normal speech signal,this paper proposes a immune optimized fast K-SVD sparse representation algorithm.Take the advantage of speech sparse representation,the high dimension of lying speech signal can be reduced,meanwhile,the sparse characteristic basis can also be extracted.Therefore,the differences between lying speech and normal speech can be analyzed by utilizing the characteristic basis.In addition,the deep learning network of Stacked Sparse Automatic Encoder(SSAE)is introduced.Deep learning can obtain person's common lying characteristics through analysis a large number of samples.As the basic acoustic feature is lack of lie expression,so the deep characteristic is a compensation for it.The experimental results show that,the correct recognition rate of individual lie has increased 4%-10% due to the introduction of deep characteristics.This suggests that the method of lie detection based on speech sparse representation is feasible.The characteristic extraction method proposed in this paper provides a new research way of psychological calculation.2.A time-series model of Dynamic Sparse Bayesian Networks(DSBN)for lying speech is proposed.Lying speech has significant temporal characteristics,but the general modeling method without giving sufficient thought of the typical time dynamic characteristics of lie,not to mention the combining with characteristics of psychological calculation.Aiming at this problem,a time-series modeling method based on DSBN for lie speech analysis is presented in this paper.This method uses the topology of DSBN to reflect the probability dependent relationship of the lying psychological state variables and the situation of all the variables changed by time.Thus,the correlation relationship and order relationship between the characteristics can be calculated.The simulation experiments shows that,the lie speech recognition can be effectively achieved by using lying state time-series model which is established in this paper.The average correct recognition rate had reached76%.Therefore,the introduced DSBN time-series model can successfully express the dynamic process of lying psychological state.In a word,this study provides an efficient time-series modeling method for lying psychological state calculation.3.A research framework of lying psychological state analysis and recognition based on sparse representation is introduced in this paper.Because the change process of lying psychological state is complexity,and which factors effect the speech signal are unpredictability.As a result,there are some limitations to study lie speech according to the experience of extracting basic acoustic characteristics.In addition,lie signal intermittently hide in the speech fragments,and it is a repeated dynamic change process.Therefore,this paper presents a lying psychological state recognition framework based on speech sparse representation.Under this research framework,the deep unsupervised learning model with sparse constraint condition is adopted to mining the main lying psychological structure characteristics from speech signals.Building a dynamic sparse bayesian time-series model for lying speech,then the lying psychological state recognition can be effectively implemented.From what has been discussed above,it can conclude that the lie detection scheme based on the analysis of speech proposed in this paper is a new exploration for psychological calculation.The research results of this paper have enriched the speech sparse representation theory,and it promoted the development of lie detection technology based on speech analysis.In a word,this research has important significance of academic exploring and practical application prospect.