Two-level recognition of isolated word using neural nets

Describes a neural-net based isolated word recogniser that has a better performance on a standard multi-speaker database than the reference hidden Markov model recogniser. The complete neural net recogniser is formed from two parts: a front-end which transforms the complex acoustic specification of the speech into a simplified phonetic feature specification, and a whole-word discriminator net. Each level was trained separately, thus considerably reducing the time necessary to train the overall system