We use a generalized classification method which is based on the thermodynamic measure of free energy, i.e. the method does not use conventional probabilities as discriminant function. This method is used for decoding in a speech recognition system based on Hidden Markov Models (HMM). The free energy classification method has the temperature T as an adjustable parameter. In this generalized scheme the values T _ 0 and T _ 1 result as special cases in the widely used Viterbi decoding and the maximum a posteriori (MAP) classification, respectively. We present an approximation of the free energy method which is easy to implement in standard speech recognition systems and allows to study this method. The HMMs are trained with a conventional speech recognition system on subsets of the Wall Street Journal corpus. As test set we use another subset with additional noise. We find that temperatures T _ 1 may yield to better recognition results of these mismatched test data than the conventional methods, showing a potential for improved recognition of noisy speech.
Cite as: Krüger, S.E., Barth, S., Katz, M., Schafföner, M., Andelic, E., Wendemuth, A. (2004) Free energy classification at various temperatures for speech recognition. Proc. 9th Conference on Speech and Computer (SPECOM 2004), 104-107
@inproceedings{kruger04_specom,
author={S. E. Krüger and S. Barth and M. Katz and M. Schafföner and E. Andelic and Andreas Wendemuth},
title={{Free energy classification at various temperatures for speech recognition}},
year=2004,
booktitle={Proc. 9th Conference on Speech and Computer (SPECOM 2004)},
pages={104--107}
}