A Preliminary Study on the Feature Distribution of Deceptive Speech Signals

A preliminary study is conducted to compare the feature distribution between normal and deceptive speech, and the results are reported in this paper. The objective of this research is to show that deceptive speech may be recognized through the acoustic parameters of general speech characteristics. Six speech parameters, i.e., Mel-frequency Cepstral Coefficients (MFCC), Relative Spectral Filter Perceptual Linear Prediction (RASTA-PLP), pitch frequency, time-domain samples, zero-crossing rate and fractal dimension are used in the statistics. The distributions of these parameters indicate clear differences between the two speech styles. The lowest average degree of difference for these features was 4.74%, and the highest degree was over 20%. Therefore, the distribution demonstrates that there is significant distinction between speech relating the truth and speech relating falsehoods. Linear Discriminant Analysis (LDA) and the Gaussian Mixture Model (GMM) are used to recognize the two psychological states of people’s pronunciation, with accuracy above 50%. The results show that there is in fact deceptive information in speech signals and that it can be detected by pattern recognition. These findings provide the theoretical basis for detecting deception in speech signals.