Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Improved Acoustics Modeling for Speech Recognition Using Transformation Techniques

Carrson C. Fung, Oscar C. Au, Wanggen Wan, Chi H. Yim, Cyan L. Keung

Human Language Technology Center, Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong

In statistical speech recognition, misclassification often occurs when there is a mismatch between the incoming signal and the acoustics model inside the recognizer. In order to combat this problem, techniques such as Cepstral Mean Subtraction, Vocal Tract Normalization, adaptation and pronunciation model can be used.

In this paper, we proposed a new approach based on transformation technique where the output distribution function in the HMM model, a Gaussian probability density function, could be transformed to match the estimated distribution of the incoming signal by using a memoryless invertible nonlinearity function. Since the new density still has a Gaussian form, the function could be completely characterized by using the Expectation Maximization (EM) algorithm.

Full Paper

Bibliographic reference.  Fung, Carrson C. / Au, Oscar C. / Wan, Wanggen / Yim, Chi H. / Keung, Cyan L. (2000): "Improved acoustics modeling for speech recognition using transformation techniques", In ICSLP-2000, vol.4, 176-179.