INVESTIGADORES
GURLEKIAN Jorge Alberto
congresos y reuniones científicas
Título:
A preliminary approach to forensic speaker recognition using phonemes
Autor/es:
UNIVASO P, GURLEKIAN, JA AND MARTÍNEZ SOLER M
Lugar:
Madrid
Reunión:
Conferencia; IberSpeech 2012; 2012
Institución organizadora:
Universidad Autónoma de Madrid
Resumen:
Abstract. Present work focus on speaker identi cation using phonemicinformation. An ASR system -based on HMMs- is employed to extractacoustic information from phonemes as speaker identity features. Maincontribution resides on proposing phoneme forced alignment which couldbe performed at the forensic environment. This approach with manualintervention, is further combined with a GMM approach, resulting in anexpected improvement on performance. Tests on a xed-phone databaseof 136 Argentine-Spanish speakers were performed to evaluate the pro-posed approach in two conditions: rst, using the same recording sessionand channel for both the suspect and evidence and second adding babblenoise to the evidence. Results for the ideal condition show an error ratereduction of 68% relative to the GMM baseline system. The di erencein favor of HMM with forced alignment against GMM holds for di erentSNR conditions.
