This paper describes a set of experiments on
training and search techniques for development of a
neural-network based continuous digits recognizer. When
the best techniques from these experiments were combined
to train a final recognizer, there was a 56% reduction in
word-level error on the continuous digits recognition task.
The best system had word accuracy of 97.67% on a test set
of the OGI 30K Numbers corpus; this corpus contains
naturally-produced continuous digit strings recorded over
telephone channels. Experiments investigated the effects
of the feature set, the amount of data used for training, the
type of context-dependent categories to be recognized, the
values for duration limits, and the type of grammar. The
experiments indicate that the grammar and duration limits
had a greater effect on recognition accuracy than the
output categories, cepstral features, or a doubling of the
amount of training data. In addition, the forwardbackward method of training neural networks was
employed in developing the final network.
