Improving Phone-level D in Lda with Subphone-l

Linear Discriminant Analysis (LDA) is a well-known technique to improve the discrimination among classes or reduce the dimensionality with minimum loss of discrimination. It is generally used as a part of the front-end of speech recognizer with the classes defined on phone or subphone level. LDA with subphone-level classes (such as tied states or individual Gaussian densities) shows superior performance to that with phone-level classes. However it does not seem to provide the best discrimination from viewpoint of recognition performance. This paper focuses on improving the discrimination between phones while subphones such as tied states are used as the basic classes of LDA. To this end, this paper proposes a method to control the discrimination ability among subphone classes so as to de-emphasize (or ignore) the discrimination between those belonging to the same phone, while emphasizing the discrimination between those belonging to the different phones. Proposed method is implemented in the framework of Weighted Pairwise Scatter LDA (WPS-LDA) and is evaluated on Korean connected digit recognition task. According to experimental results, proposed method shows higher performance than conventional LDA and WPS-LDA. Performance is the highest when the discrimination between pairwise classes belonging to the same phone is totally ignored. The string error reduction rates of conventional LDA, WPS-LDA and proposed method are 13%, 14.5% and 23%, respectively.