From 'past history' to 'interactive activation' in speech recognition

An alternative to the strictly sequential approach to the speech signal and its processing and recognition is proposed. Embodying parallel processing and associative modelling, it bears many parallels to current interactive activation and distributed memory models which have recently been receiving renewed interest, particularly in the perception of visual stimuli. It is suggested here that one characte r istic of such models is that they deal with a pattern of activation within the total system at any moment in time, and not with the past sequence of stimulus events which has resulted in this activation . A computer simulation for speech recognition based on such prinCiples has been extended to a vocabulary of over 100 words, and demonstrates both the power of such an approach, and some of its limitations when applied only at a single local level of stimulus information.