Some experiments in spoken language acquisition

The authors report on some rudimentary experiments involving speech understanding devices which adaptively acquire the language for this task. They focus on a particular technique issue which arises in such spoken language acquisition devices, namely that the device needs to decide whether an utterance is a new word or a perturbation of some known word. Such systems have two unique characteristics. First, no text is provided to the device, in contrast to all other existing speech understanding systems. Second, the vocabulary and grammer are unconstrained, being acquired by the device during the course of performing its task. This is also in contrast to all other systems, where the salient vocabulary and grammar are preprogrammed. The experimental mechanism is a connectionist network embedded in a feedback control system.<>