Manifold HLDA and its application to robust speech recognition

A manifold heteroscedastic linear discriminant analysis (MHLDA) which removes environmental information explicitly from the useful information for discrimination is proposed. Usually, a feature parameter used in pattern recognition involves categorical information and also environmental information. A well-known HLDA tries to extract useful information (UT) to represent categorical information from the feature parameter. However, environmental information is still remained in the UI parameters extracted by HLDA, and it causes slight degradation in performance. This is because HLDA does not handle the environmental information explicitly. The proposed MHLDA also tries to extract UI like HLDA, but it handles environmental information explicitly. This handling makes MHLDA-based UI parameter less influenced of environment. However, as compensation, in MHLDA, the categorical information is little bit destroyed. In this paper, we try to combine HLDA-based UI and MHLDA-based UI for pattern recognition, and draw benefit of both parameters. Experimental results show the effectiveness of this combining method.