A continuous unsupervised adaptation method for speaker verification

Abstract : —This paper deals with unsupervised model adaptation for speaker verification. We proposed a new method for updating speaker models using all test information incoming in the system. This is a continuous adaptation method which relies on the probability of the test trial belonging to the target speaker. Our adaptation scheme is evaluated in the framework of the NIST SRE 2005. This approach reaches a relative improvement for the NIST unsupervised adaptation mode of 15% DCF and 35% EER. I. INTRODUCTION Gaussian Mixture Models (GMM) based systems for speaker recognition are widely used in speaker recognition applications due to their robust performance [1]. Despite new normalization techniques such as Bayesian Factor analysis [2], GMM based speaker verification systems show their limits when limited enrollment data are available. A way of passing through this problem is to increase this amount of data by using test information [3,4,5]. Indeed during the normal use of the verification system some test trials are true target trials and could be added to the enrollment data. This is known as unsupervised adaptation.
Document type :
Journal articles
Complete list of metadatas

https://hal.archives-ouvertes.fr/hal-01311567
Contributor : Bibliothèque Universitaire Déposants Hal-Avignon <>
Submitted on : Wednesday, May 4, 2016 - 2:11:15 PM
Last modification on : Saturday, June 15, 2019 - 12:24:17 PM

Identifiers

  • HAL Id : hal-01311567, version 1

Collections

Citation

Alexandre Preti, Jean-François Bonastre, Francois Capman. A continuous unsupervised adaptation method for speaker verification. Innovations in E-learning, Instruction Technology, Assessment, and Engineering Education, 2007. ⟨hal-01311567⟩

Share

Metrics

Record views

71