Representation, scaling, and time invariance in neural network speech recognition: evidence for the recognition of stop consonants and vowels

Summary form only given, as follows. A neural network speech recognition system was constructed based on a feedback network with backpropagation in an effort to explore various issues relating to representation, scaling, generalization, and time invariance. The network was trained on stop consonant and vowel data obtained from continuous speech. The results indicated a surprising tendency for the network to construct local rather than distributed representations. The stability of learning increased with network size, and generalization was not impaired in large scale networks. Increased complexity of the recognition problem did reduce recognition performance and learning stability, and increase learning time. The network showed capabilities of learning to recognize speech sounds that were not synchronized to a particular time.<>