Until now we concentrated our work on methodological questions. Here we report on experimental results which show that recurrent neural networks being trained to behave as phoneme spotters can be used to transform complex speech patterns into simpler and more unique sequences of events. They are able to deal with the fairly different time structure of the different phoneme classes. The representations stored of at least plosives become more robust, if the training covers several speakers, since the networks are forced to extract only the relevant information.
Bibliographic reference. Wittenburg, P. / Couwenberg, R. (1991): "Recurrent neural nets as building blocks for human word recognition", In EUROSPEECH-1991, 1015-1018.