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

Analytic Methods for Acoustic Model Adaptation: A Review

Shigeki Sagayama1 (,2,) Koichi Shinoda (3), Mitsuru Nakai (2) and Hiroshi Shimodaira (2)

(1) The University of Tokyo, Tokyo, Japan
(2) Japan Advanced Institute of Science and Technology, Tatsu-no-kuchi, Ishikawa, Japan
(3) NEC Laboratories, Miyazaki, Miyamae-ku, Japan

This paper discusses analytic methods of acoustic model adaptation for automatic speech recognition and reviews other major methods. The main purpose of this paper is to demonstrate the potential of analytic approach for model adaptation. As an example of analytic methods, Jacobian Adaptation (JA) is intensively discussed and its potential of applicability to speech recognition problems is revealed. Vector Field Smoothing (VFS) is introduced as an extension of a special case of JA. Other method reviewed in this paper include Maximum A Posteriori (MAP) estimation, transformation-based approaches including Maximum Likelihood Linear Regression (MLLR), structural approaches, model selection including Eigenvoice, and feature compensation including Speaker Adaptive Training (SAT).

Full Paper

Bibliographic reference.  Sagayama, Shigeki / Shinoda, Koichi / Nakai, Mitsuru / Shimodaira, Hiroshi (2001): "Analytic Methods for acoustic model adaptation: A review", Invited Lecture, In Adaptation-2001, 67-76.