Experiments with artificial neural networks for phoneme and word recognit
Kjell O. E. Elenius, Mats Blomberg · 1993
An artificial neural network has been bained by the error back-propagation technique to recopse phonemes and words. The speech material was recorded by a male Swedish talker and was labelled by a phonetician. There were 38ou put nodes corresponding to Swedish phonemes. Introducing coarticulation information by adding simple recurrency to the net is shown to be more effective than expanding the size of the input spectral window. The phoneme recognition network was used with dynamic programming for time alignment to recognise connected digits in a speakerindependent way. It was compared to a similar recogniser based on nine quasi-phonetic features instead of 38phonemes. The phoneme-based system performed better fhan the feature-based one for five out of seven speakers.