Sixth European Conference on Speech Communication and Technology
(EUROSPEECH'99)

Budapest, Hungary
September 5-9, 1999

A Segmental Approach to Text-Independent Speaker Verification

J. Cernocky (1), D. Petrovska-Delacrélaz (2), S. Pigeon (3), P. Verlinde (3,4), Gérard Chollet (4)

(1) Brno University of Technology, Inst. of Radioelectronics, Czech Republic
(2) EPFL Lausanne, DE-CIRC, Switzerland
(3) Royal Military Academy Brussels, SIC, Belgium,
(4) ENST Paris, Département Signal, France

Current text-independent speaker verification systems are usually based on modeling globally the probability density function (PDF) of the speaker feature vectors. In this paper, segmental approaches to text-independent speaker verification are discussed. Unlike the schemes based on Large Vocabulary Continuous Speech Recognition (LVCSR) with previously trained phone models, our systems are based on units derived in unsupervised manner using the ALISP (Automatic Language Independent Processing) tools. Speaker modeling is then done independently for each class of speech sounds. Among the techniques to merge the classdependent scores, linear combination was tested and logistic regression and a method based on the Mixture of Experts technique are under investigation. The experimental results were obtained on the data from the NIST-NSA'98 campaign.

Keywords: text-independent speaker verification, segmental approach, data fusion.


Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Cernocky, J. / Petrovska-Delacrélaz, D. / Pigeon, S. / Verlinde, P. / Chollet, Gérard (1999): "A segmental approach to text-independent speaker verification", In EUROSPEECH'99, 2207-2210.