Self-Organized Continuous Speech Recognition

Abstract

Current efforts to recognize continuous (or “connected”) speech are aimed at constructing a voice-excited “typewriter” that automatically transcribes natural speech into ordinary (e.g. English) written form. So far, however, only very restricted speech has been recognized. The sentences that are spoken must either be prescribed a priori by an artificial grammar which the experimenter has designed, or else limited by a vocabulary and a restricted area of discourse such as that used in business letters, book reviews, or airline reservation systems. These latter so-called natural tasks are generally much more difficult than the artificial ones (given a fixed vocabulary).

At IBM-Research we are now working on the recognition of continuous natural speech, without the aid of artificial pauses between words, or unnaturally small vocabularies or artificial constraints on syntax. As of this writing ours is the only place in the United States (and, as far as we know, the world) with this goal. The particular utterances to be recognized are chosen from the Laser Patent Text, consisting of patent disclosures involving laser devices. Previously (1978) we succeeded in recognizing perfectly sentences generated by an artificial grammar known as New Raleigh.