Dynamic methods in speech processing

Abstract

Many experiments in phonetics and other areas of speech science depend on new speech-processing techniques for extracting parameters to be used in the experiments. An example is LPC analysis, and resynthesis with changed duration, pitch, or some other parameter. Parameters are usually estimated in a static, context-free manner; in some cases dynamic methods lead to improvement and/or simplification of the process. Some examples include: (a) Making an automatic voiced-unvoiced decision (as an aid in doing phonetic marking) based on context; (b) finding “pitch pulses” (points of glottal closure) with a simple form of Dynamic Programming (sometimes called Viterbi's algorithm); and (c) time-aligning two LPC-resynthesized tokens of an utterance (warping one to fit the other) using Bridle's algorithm, another form of Dynamic Program. These and other examples will be discussed and illustrated.