Feature Space Reduction and Decorrelation in a Large Number of Speech Recognition Experiments

J.V. Psutka, L. Mller, L. Šmdl, and J. Psutka (Czech Republic)


Automatic speech recognition, feature space reduction, PLP-based parameterization


The paper studies the influence of significant space reduction and decorrelation techniques on the performance of an automatic speech recognition (ASR) system. A baseline PLP-based ASR system, which works with zero-gram language model, was trained using speech of one thousand speakers. The Linear Discriminant Analysis (LDA), Heteroscedastic LDA (HLDA) and Smoothed HLDA (SHLDA) techniques were applied and compared in many experiments in which an optimum dimension of a feature space was searched for.

Important Links:

Go Back