Source Digital Voice Recorder Identification by Wavelet Analysis

Abstract

Identification of the speech signal origin is an important issue since it may play a vital role for criminal and forensic investigations. Yet, in the media forensics field, source digital voice recorder (DVR) identification has not been given much attention. In this paper we study the effect of subband based features obtained using uniform wavelet packet decomposition and Teager energy operator on the DVR model and brand identification performance. In order to assess the effects of these features on the proposed system, one-class classifiers (OCCs) with two reference multi-class classifiers were carried out. The performance of the DVR identification system is tested on a custom database of twelve portable DVRs of six different brands. The results showed that the proposed system can effectively identify the correct DVR brands/models with a high accuracy. Moreover, it was observed that the combination of the traditional speech features with subband Teager energy cepstral parameters (STEC) and short time frame energy as a feature improved recognition accuracy under both silent and noisy recording conditions.