Influence of Natural Voice Disguise Techniques on Automatic Speaker Recognition

The problem of voice disguise is usually investigated in the context of surveillance and forensics. The biometric techniques applied for automatic or subj ective speaker recognition can be deliberately or non-deliberately misled by technical or natural methods. The investigations presented in this paper include data collection and automatic speaker recognition scores. The database consists of the utterances of several natural voice disguise techniques: phonation (raised and lowered pitch, whisper), phonemic (foreign accent), prosodic (speech tempo) and deformation (pinched nostrils and clenched jaws). Speaker verification was carried out with the state-of-the-art system of MFCC (Mel Frequency Cepstral Coefficients) feature extraction and GMM (Gaussian Mixture Models) classification.