Training methods for a connectionist model of consonant-vowel syllable recognition

A description is given of several CV (consonant-vowel) syllable recognition experiments using neural network learning and retrieval paradigms. Previously (1988), the authors trained both one- and two-speaker systems to classify phonemes from the set (b d g)*(a i u) at a rate over 90%. However, when more phonemes are added to the system, additional training techniques are necessary to maintain this system performance. These techniques include training parallel sets of hidden units sequentially. In this technique the weights associated with one set are frozen while the weights of the other set are being trained. In addition, it was found that elaboration of the same input data in alternate representations helps performance considerably. Finally, it is shown that training the system to reject nonspeech acoustic data helps system performance when error analysis criteria are made more rigorous to test for practical speech recognition performance.<>