Linguistically informed automatic speech recognition

Abstract

The development of an automatic event-based speech recognition system (EBS) that relies heavily on acoustic phonetics (to guide the recognition process and to extract relevant information) and combines a phonetic-feature hierarchy with a uniform statistical framework (at present, Support Vector Machines) to provide adaptability and flexibility is currently under way. This recognition framework allows for easy assessment and distinction of the performances of the acoustic parameters versus that of the pattern recognizer. The overall structure of EBS involves (1) landmark detection based on acoustic parameters that are related to the source and manner-of-articulation phonetic features and (2) use of the landmarks in the extraction of other acoustic parameters related to the place-of-articulation phonetic features. This talk will focus on the development of the acoustic parameters and the need for relative parameters for speaker independence, multi-time-scale processing to capture the dynamics of phonetic segments and extensive evaluation of the parameters to hone in on direct measures of the relevant acoustic properties. [Work supported by NSF and NIH.]