Using contextual information of phones is an effective way to improve the performance of phone classification tasks, but requires the use of dimensionality reduction. One of the disadvantages of Linear Discriminant Analysis (LDA), a popular dimensionality reduction method is that it is not able to account for local differences between the distributions of classes in the feature space. Newer methods, such as the Local Fisher Discriminant Analysis (LFDA), on the other hand, may overestimate the contribution of local distributions. In this paper, we propose to use a dimensionality reduction algorithm with an affinity matrix that allows finding the optimal trade-off between local and global information. Experiments on TIMIT show that both local and global information in the MFCC feature space are important for phone classification and that a substantial improvement can be achieved over both LDA and LFDA.
展开▼
