A linear transformation is described that turns a set of variables into a new set all having the same variance. The number of new variables is at least that of the original set and optionally more. This reversible transformation can be used to encode data efficiently when the precision available is limited and is the same for all variables. In tests with synthetic data, it encoded information more efficiently than simple scaling methods. Also, because the transformation is orthonormal, it can be used directly in speech recognition systems that use unweighted Euclidean distances. In speaker-independent alphabet recognition tests with telephone speech, scaling the variables into four bits caused a 34% increase in error rate. Applying the transformation before scaling, however, resulted in only a 5% rise. Finally, spreading the information in the original 16 variables over 32 transformed variables before scaling resulted in an increase of just 3%.
Cite as: Hunt, M.J. (1995) The even transform: a variance-equalizing orthogonal transformation and its application to speech recognition. Proc. 4th European Conference on Speech Communication and Technology (Eurospeech 1995), 931-934, doi: 10.21437/Eurospeech.1995-216
@inproceedings{hunt95_eurospeech,
author={Melvyn J. Hunt},
title={{The even transform: a variance-equalizing orthogonal transformation and its application to speech recognition}},
year=1995,
booktitle={Proc. 4th European Conference on Speech Communication and Technology (Eurospeech 1995)},
pages={931--934},
doi={10.21437/Eurospeech.1995-216}
}