Parser for Feature - Based Speech Recognition

Many phonemic units can be readily described as a configuration of segments of acoustic features. Certain stop allophones, for example, can be constructed out of the features silence, hurst release and aspiration. It seems natural to specify this as a rewrite rule that can be interpreted by a parser: stop <silence + burst + aspi ration. Conventional parsers construct grammatical phrases out of words. The words and phrases are discrete and follow each other in a well-defined order. Suchparsersare not appropriate for acoustic segments because these are variable in length, and overlap each other or have gaps separating them. This paper describes the capabilities of a parser that has been modified to handle this kind ofinput.