Reasoning About the Acoustic Realisation of Semivowels Using an Intermediate Representation - The 'Speech Sketch'

We argue for an approach to acoustic phonetic reasoning in speech recognition based on the 'Speech Sketch'a data structure in which the behaviour of spectral parameters is explicitly described.We represent acoustic-phonetic knowledge in frame-like structures whose terminals map onto the descriptors of the speech sketch.Part of a frame expresses how the manifestation of the event the frame describes is expected to change with context.The speech sketch describes what has happened and the frames describe, in an equivalent way.what is expected to happen in particular speech events.'Recognition' within this paradigm, is achieved by a recursive matching function which resembles the 'inference engine' of an expert system.We discuss first results from a testbed implementation of the above ideas, in which a speech sketch for formant frequency data is used to distinguish between /1/, /r/, [w/ 5 /j/ in various environments.