A supervised two-channel learning method for hidden Markov model and application on lip reading

In this paper, a novel two-channel learning method for hidden Markov model (HMM) is proposed. This method is specially designed to train HMMs for fine recognition from similar observations. The prominent features of this method are 1.) the criterion function is based on the difference between trai...

Full description

Saved in:
Bibliographic Details
Main Authors: Foo, Say Wei, Dong, Liang
Other Authors: IEEE International Conference on Advanced Learning Technologies (2nd : 2002 : Kazan, Russia)
Format: Conference or Workshop Item
Language:English
Published: 2009
Online Access:https://hdl.handle.net/10356/90829
http://hdl.handle.net/10220/4617
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:In this paper, a novel two-channel learning method for hidden Markov model (HMM) is proposed. This method is specially designed to train HMMs for fine recognition from similar observations. The prominent features of this method are 1.) the criterion function is based on the difference between training sequences, and 2.) a twochannel structure is adopted to maintain the validity of the HMM. This learning method has been applied on a viseme-level lip reading system. The result shows that the performance of the two channel approach is better than that of the maximum likelihood (ML) estimation.