Workshop on the Auditory Basis of Speech Perception Keele University, UK July 15-19, 1996

Recognising Occluded Speech

Martin Cooke, Andrew Morris, Phil Green

Department of Computer Science, University of Sheffield, Sheffield, UK

Auditory representations of clean speech contain much redundancy. Arguably, it is this redundancy which enables listeners to recognise speech in adverse conditions. Under the assumption that some time-frequency regions are too heavily masked to derive any useful estimate of speech level, the auditory system faces the missing data problem: how to achieve robust performance with incomplete evidence. This paper develops several techniques for overcoming the missing data problem and demonstrates that high-performance automatic speech recognition can be achieved for quite high degrees of data masking, but that further information, such as provided by context or through auditory induction constraints, will be required to handle realistic masking conditions.

Full Paper

Bibliographic reference.  Cooke, Martin / Morris, Andrew / Green, Phil (1996): "Recognising occluded speech", In ABSP-1996, 297-300.