Speech forensics: Automatic acquisition device identification.
Daniel Garcia-Romero, Carol Espy-Wilson · The Journal of the Acoustical Society of America · 2010
Automatic identification-of-the-source is an important problem in the field of speech forensics. That is, given a speech recording, the problem is to determine how to automatically determine the exact make and model of the acquisition device regardless of the speaker and speech content. Recent work developed by the authors have shown that statistical modeling of the device characteristics in terms of MFCCs provides excellent discriminative power when the recordings are obtained under controlled conditions (i.e., all recordings come from the same room and a unique device per class is used). In this study we propose to expand these results by analyzing the effects that room acoustics and intra-device variability have in the source identification accuracy. To perform this analysis, the NIST 2008 SRE dataset will be used. This corpus provides recordings acquired with multiple devices of the same make and model in different rooms.