ISCA Archive Eurospeech 1993
ISCA Archive Eurospeech 1993

Word and phrase spotting with limited training

M. J. O'Kane, P. E. Kenne

Experiments using a computationally efficient wordspotter which works well on very limited training data are described. Most wordspotters reported in the literature are either HMM or neural network based. This wordspotter uses yet another architecture based on parallel broad encodings. The most interesting aspect of the use of this wordspotter is the results it achieves with limited training data even from a single speaker. Improvements in wordspotting as the amount of training data is increased are given as a function of the speakers in the training set and the amount of training data for each speaker. Speaker-dependent results are also given.

Keywords: wordspotter, training