This paper describes a set of experiments on neural-network
training and search techniques that, when combined, have
resulted in a 54% reduction in error on the continuous digits
recognition task. The best system had word-level accuracy of
97.52% on a test set of the OGI 30K Numbers corpus, which
contains
naturally-produced continuous digit strings recorded
over telephone channels. Experiments investigated 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 50% increase in the amount of training
data.
