Speech Deception Detection Based On Wavelet Packet Transform And Sparse Decomposition

Deception detection has important and widespread applications in public places security inspection,employment career evaluation,police investigation and insurance financial credit assessment,also has aroused many attentions in academia.The speech signal is more convenient and natural than the EEG signals,heartbeat and respiration rate,skin sweat volume and polygraphs,which rely on human physiological signals and professional instruments,which can not cause contradictive,fear and stress emotion of subject.Speech deception detection research involves feature extraction,signal analysis,classification and identification,is still in the exploratory stage.How to use speech features to express deceptions information,and the establishment of appropriate models to detect deceptions are to be studied in depth.The method of speech deception detection feature extracting with wavelet packet transform and sparse decomposition algorithm is proposed in is paper,also some machine learning modules are used for constructing deception detection system.The main research work is as follows:Firstly,we expound the research background and significance of speech deception detection,analysis the research history of deception detection and the internal and abroad research status of speech deception detection,and then indicate the main problems in deception detection research.Secondly,the wavelet packet band cepstral coefficient is proposed as a speech feature for deception detection.Due to the current deception detection research is less focus on the frequency domain characteristics of speech signal.The wavelet packet which has good analysis characteristics of time-frequency and multi-level subdivision of speech signal frequency band,that is considered to be used with cepstral computation for capturing middle and high frequency band details about deception as the wavelet packet band cepstral coefficient.Then a series of comparative recognition experiment is processed with some classifier,the optimal signal frequency band division scheme is determined by the recognition results.The effectiveness of the wavelet packet band cepstral coefficient feature is verified.Finally,the sparse decomposition algorithm is proposed to extract sparse feature of multi-combined feature,and support vector machine is used to realize deception detection.In order to decrease redundancy data and high dimensions in multi-combined feature set,sparse decomposition algorithm that combines dictionary learning with sparse coding,sparsely decomposes the set of fusion features to obtain sparse features.Through a large number of comparative experiments,the method of obtaining the optimal over-complete dictionary for sparse representation and the dimension of sparse features are determined.The experiment results show that,the sparse cepstral feature which is extracted by combining the wavelet packet band cepstral coefficient and the Mel frequency spectral coefficient,obtains the better deception detection results under the support vector machine classifier,improve the performance of the deception detection system significantly.Furthermore,the performance of the proposed algorithm of sparse decomposition for feature is proved to be better than principal component analysis and K-singular value decomposition dictionary learning algorithm.