Blind Detection of Electronic Voice Transformation with Natural Disguise

摘要

Electronic voice transformation with natural disguise ability is a common operation to change a person's voice and conceal his or her identity, which can easily cheat human ears and automatic speaker recog-nition(ASR) systems and thus presents threaten to security. Till now, few efforts have been reported on detection of electronic transformation, which aims to distinguish disguised voices from original voices. Therefore in this paper we investigate the principle of electronic voice transformation, and propose a blind detection approach using MFCC(Mel Frequency Cepstrum Coefficients) as the acoustic features and VQ-SVM (Vector Quantization-Support Vector Machine) as the classification method. By extensive experiments, it is demonstrated to have classification accuracy higher than 98% in most cases, indicating that the proposed approach has good performance and can be used in forensic applications.