Phoneme recognition with neural networks using a novel fuzzy training algorithm

A fuzzy training algorithm is proposed to improve the phoneme recognition performance of neural networks (NNs) as an alternative to the conventional back-propagation training algorithm with hard-decision supervision (HDS). The conventional algorithm uses a hard decision criterion (one and zero supervisor signals) and suffers from the overlearning problem. In contrast, the proposed training algorithm uses fuzzy-decision supervision which has a gray scale between zero and one. In phoneme sample space, the algorithm uses neighboring information or considers overlap regions of phoneme distribution boundaries. The supervisor signal for a training sample is determined reflecting the distribution of the other training samples around the original sample. It therefore softens the hard decision criterion for phoneme recognition and prevents the overlearning problem. Several NN architectures are trained with the fuzzy training algorithm by using b, d, g, m, n and N from the Japanese vocabulary uttered in different speaking styles. The proposed algorithm is shown to improve the performance over the conventional HDS algorithm