Fuzzy partition models and their effect in continuous speech recognition

Abstract

Fuzzy partition models (FPMs) with multiple input-output units were applied to continuous speech recognition, and the use of automatic incremental training was evaluated. After initial training using word data, phrase recognition rates of 72.7% and 66.9% were obtained for an FPM and a TDNN (time-delay neural network), respectively. After incremental training, the phrase recognition rates improved to 86.3% and 78.4%, respectively. The FPMs provided more accurate segmentation after incremental training. The experiments determined that better phoneme segmentation provides greater improvement in phrase recognition. Incremental training also significantly improves recognition performance. As FPMs can be trained rapidly, various applications using large-scale training data are also possible.<>