Alpha-nets interpret a set of Hidden Markov Models (HMMs) as a connectionist network enabling back propagation to be used to adapt the parameters of the HMMs, thus providing a powerful technique for improving the discriminatory ability of the HMMs. Previous work has focussed on the adaption of the state parameters of models in which the emission probability is described by a single mode Gaussian distribution. The parameters adapted can be the mixture coefficients themselves in addition to the multivariate means and variances or the elements of the input transformation. The input transformation may be linear or non-linear. We provide experimental results which show that, in a speaker independent isolated digit recognition task using telephone quality speech, adapting the parameters of the input transformation can reduce the error rate observed on the training set, while testing with unseen material produces less improvement in the system performance.
Cite as: Carey, M.J., Parris, E.S. (1991) Adapting input transformations using alpha-nets for whole word speech recognition. Proc. 2nd European Conference on Speech Communication and Technology (Eurospeech 1991), 555-558, doi: 10.21437/Eurospeech.1991-138
@inproceedings{carey91_eurospeech,
author={Michael J. Carey and Eluned S. Parris},
title={{Adapting input transformations using alpha-nets for whole word speech recognition}},
year=1991,
booktitle={Proc. 2nd European Conference on Speech Communication and Technology (Eurospeech 1991)},
pages={555--558},
doi={10.21437/Eurospeech.1991-138}
}