Blind speaker identification for audio forensic purposes

This paper presents a blind method for speaker identification for audio forensics purposes. It is based on a decision system with fuzzy rules and works with the correlation between the cochleagrams of the audio proof and of the audios of the suspects. Our proposed system can give a null output, a unique selected suspect or a group of identified suspects. According to several tests, our Overall Accuracy (OA) is 0.97 with agreement (κappa index, κ) of 0.75. Additionally, unlike typical systems in which a low false acceptance (FP) implies high false rejection (FN), our system can work simultaneously with FN and FP equal to zero (i.e. OA=1; κ=1). Finally, our system works with blind identification, it means, without preliminary knowledge of the audio recordings or a training step; an imperative characteristic for audio forensics.