EUROSPEECH 2003  INTERSPEECH 2003

This work discusses the improvements which can be expected when applying linear featurespace transformations based on Linear Discriminant Analysis (LDA) within automatic speechrecognition (ASR). It is shown that different factors influence the effectiveness of LDAtransformations. Most importantly, increasing the number of LDAclasses by using timealigned states of HiddenMarkovModels instead of phonemes is necessary to obtain improvements predictably. An extension of LDA is presented, which utilises the elementary Gaussian components of the mixture probabilitydensity functions of the HiddenMarkovModels' states to define actual Gaussian LDAclasses. Experimental results on the TIMIT and WSJCAM0 recognition task are given, where relative improvements of the errorrate of 3.2% and 3.9%, respectively, were obtained.
Bibliographic reference. Schaffoner, M. / Katz, M. / Kruger, S.E. / Wendemuth, A. (2003): "Improved robustness of automatic speech recognition using a new class definition in linear discriminant analysis", In EUROSPEECH2003, 28412844.