A Prototype Speech Recognizer based on Associative Learning and Nonlinear Speech Analysis

The grouping of sounds has been shown to be partially based on amplitude modulation (AM) characteristics (Bregman et al., 1985), suggesting that AM information observed in auditory nerve fibers could be used by the auditory system to segregate speech from background noise. This chapter proposes a speech recognizer prototype that relies on patterns of modulation observed in auditory fibers influenced by a summation of close harmonics. The recognition task is performed by a modified Dystal (DYnamically STable Associative Learning) neural network (Alkon et al., 1990) (Blackwell et al., 1992). Preliminary results indicate that the approach might be efficient and powerful. Further experiments have still to be done in order to evaluate the approach. Experiments on continuous speech in noisy environment are planned.