ITRW on Adaptation Methods for Speech Recognition August 29-30, 2001 Sophia Antipolis, France

Comparison of Techniques for Environment/Application Adaptation in a Telephony Context

Lionel Delphin-Poulat

France Télécom R&D DIH/IPS, Lannion, France

Model adaptation is a key point to have reliable speech recognizers for practical applications. In this article, we recall the main adaptation techniques: Maximum a Posteriori and MLLR Adaptation. We emphasize the sources of knowledge they use and show how to take them into account; for MLLR a data-driven clustering procedure is presented: a Gaussian tree is built at training time; based on this tree and adaptation data, the clusters for adaptation are chosen. We report experiments on field data that show the efficiency of adaptation techniques both in supervised and unsupervised mode in practical telephony applications.

Full Paper

Bibliographic reference.  Delphin-Poulat, Lionel (2001): "Comparison of techniques for environment/application adaptation in a telephony context ", In Adaptation-2001, 139-142.