A procedure for adaptive control of the interaction between acoustic classification and linguistic decoding in automatic recognition of continuous speech

Abstract

An adaptive-control procedure is described which is intended to improve both acoustic analysis and linguistic decoding in automatic recognition of continuous speech by bringing into agreement data available at each of these stages. Specifically, hypotheses are formed by the decoder concerning the phonetic transcription derived during acoustic analysis. The procedure then accesses and utilizes relevant acoustic data in an attempt to verify or reject these hypotheses. Depending on the success of such attempts, actions are taken to constrain the decoding in subsequent processing iterations. Preliminary results are presented and discussed.