A. L. Gorin
S. E. Levinson
AT&T Bell Labs., Rm. 2D-449, Murray Hill, NJ 07974
At present, automatic speech recognition technology is based upon constructing models of the various levels of linguistic structure assumed to compose spoken language. These models are either constructed manually or automatically trained by example. A major impediment is the cost or even the feasibility of producing models of sufficient fidelity to enable the desired level of performance. The proposed alternative is to build a device capable of acquiring the necessary linguistic skills during the course of performing its task. The mechanisms underlying this research program comprise an information-theoretic connectionist network embedded in a feedback control system, and involve both spoken and typed language. Recent experiments in spoken language acquisition will be described, within the application scenario of an Inward Call Manager. This speech understanding system is unique in that its vocabulary and grammar are unconstrained, being acquired during the course of performing its task. It is also unique in that no text is involved in the training of this system. Previously, results have been reported for this system with text input, and some preliminary experiments in speech involving a small (100 utterance) database. These current experiments are more detailed, and involve a larger database of 1105 utterances.