Use of attributed grammars in speech signal processing

This paper describes a theoretical model for syntactic signal processing, and the algorithmic implementation of the model for pitch and formant tracking applications. Classical and adaptive estimation models operate on the data with a specified window function, and no use is made of structural relationships beyond adjacent frames of data. This new model consists of a grammar which accepts new data based on the statistics of the data attributes, as well as structual constraints of the grammar. High probability sentences are constructed through a nondeterministic parsing of the data. In this manner it is possible to use all of the context surrounding a particular set of windowed data to improve the estimate and segmentation is simultaneously achieved. The model is extendable to higher level constructs.