Subjective tests of speaker recognition for selected voice disguise techniques

Abstract

Research work on the effectiveness of voice disguise techniques is important for the development of biometric systems (surveillance) as well as phonoscopic research (forensics). A speaker recognition system or a listener can be deliberately or non-deliberately misled by technical or natural methods. It is important to determine the impact of these techniques on both automatic systems and live listeners. This paper presents the results of listening tests conducted on a group of 40 people. The effectiveness of speaker recognition was investigated using selected natural (chosen from four groups of deliberate natural techniques: phonation, phonemic, prosodic and deformation) and technical (pitch shifting, GSM coding) voice disguise techniques. The results were related to the previously obtained outcomes for the automatic method of verification carried out using a classical speaker recognition system based on MFCC (Mel Frequency Cepstral Coefficients) parameterisation and GMM (Gaussian Mixture Models) classification.