Pilot project on speech recognition using layered abduction and multiple knowledge types

Abstract

Our long-term goal is to test the feasibility of basing a recognition system on a model of human speech processing. Key hypotheses are: (1) Computational models for reasoning from incomplete knowledge provide a useful metaphor for many aspects of human speech understanding even at levels where highly automatic perceptual processes are at work; (2) A computational model of human cochlear processing must be used as the signal-processing front end; (3) Some representation of articulation should mediate between the acoustics and the phonology in order to accommodate contextual variation of various sorts; (4) The phonological representation must encode the prosodic structure and intonation pattern as well as the phoneme string. Accordingly our specific short-term goal is to build a small prototype system that uses a layered-abduction architecture, in which there are many stages of processing, corresponding to different levels of knowledge. The common information-processing task at each stage is to form a coherent, composite (multi-part) hypothesis that explains the data presented from the preceding levels. The input signal will be the digitized speech processed by Patterson's Stabilized Auditory Image system. An articulatory representation based on Browman and Goldstein's gestural score will mediate between the auditory representation and the phonology, and Pierrehumbert and Beckman's prosodic tree and tone string will be used for the phonological organization and intonational melody.