A preliminary approach to forensic speaker recognition using phonemes

PEDRO UNIVASO, MIGUEL MARTÍNEZ SOLER, DIEGO A. EVIN, JORGE A. GURLEKIAN

Madrid

Congreso; IberSpeech 2012; 2012

Universidad Autónoma de Madrid

Present work focus on speaker identification using phonemic information. An ASR system -based on HMMs- is employed to extract acoustic information from phonemes as speaker identity features. Main contribution resides on proposing phoneme forced alignment which could be performed at the forensic environment. This approach with manual intervention, is further combined with a GMM approach, resulting in an expected improvement on performance. Tests on a fixed-phone database of 136 Argentine-Spanish speakers were performed to evaluate the proposed approach in two conditions: first, using the same recording session and channel for both the suspect and evidence and second adding babblenoise to the evidence. Results for the ideal condition show an error rate reduction of 68% relative to the GMM baseline system. The difference in favor of HMM with forced alignment against GMM holds for different SNR conditions.