Becars: an Automatic Speaker Verification system Chafic Mokbel(1) Raphaël Blouet(2) Eduardo Sanchez Soto(2) Gérard Chollet(2) {blouet,chollet,esanchez}@tsi.enst.fr chafic.mokbel@balamand.edu.lb University of Balamand, El Koura, Libanon (2) École Nationale Supérieure des Télécommunications, Paris, France
Outline Context and aim of the work Automatic identification Biometric identification Automatic Speaker Verification Applications State of the art – research directions Conclusion / discussion
Context and aim of the work Collaboration between the University of Balamamand and l’ENST in the context of the Cedar project a French-Libanese cooperation framework Setting up and distribution of a gnu software for Automatic speaker Verification Participation of both sites to the NIST 2004 speaker verification evaluation Standard in evaluation of state-of-the-art performances
Automatic identification Why ? Secure access to sensitive data, offices or services Automatic adaptation of softwares or services to clients… How ? [S. Liu and M. Silverman 2001.] “Something that you have” : keys, cards… “Something that you know” : PIN code, password… “Something that you are” : Biometric Authentication
Biometric identification Something that you know Something that you have Biometry: something that you are SECURED SPACE Bla-bla
Automatic Speaker Verification Verification System Claimed Identity Acceptation Rejection Speech processing Biometric Technology
State of the art – research directions (1) Hypothesis Testing: Y, Acoustic Parameters Space: Mel Filter Bank Cepstal Coefficient (mfcc). Speaker Modelization : Gaussian Mixture Model (GMM), [Reynolds, 1994] Acceptation Rejection
Gaussians Mixture Model Parameters :
State of the art – research directions (3) world model, speaker independent, train with all available speaker, using the algorithm EM . client model, Obtained as an adaptation of , MAP with a prior distribution MLLR with a transform function Unified approach
Adaptation Degré de liberté variable Partitionnement variable des distributions Après chaque étape E de l’EM partitionnement donnant une quantité de données suffisante par classe 12 9 17 6 23 21 33 56
Hierarchical - MLLR adapted System
Conclusion / discussion Becars provides a state-of –the-art software for GMM parameters estimation Speed up development and research purposes of new site wishing to participate to the NIST evaluation Enhance software quality by users’ feedback