A technique termed pole focusing is presented that provides anovel approach to obtained high-resolution formant data for patternrecognition processing of the short-time speech spectrum. The techniqueoffers reliable detection of weak nasal formants and formants undergoingrapid transitions in frequency, areas where parametric spectral analysistypically performs poorly. Much of the recognition of speech using afeature-based approach relies heavily on the detection of formanttime-frequency trajectory patterns, which gives the identification notonly for the voiced speech sound currently under analysis, but also canprovide important cues to pre- and postvocalic speech. The enhancedformant detection properties offered by pole focusing therefore canconsiderably improve the reliability of formant patternrecognition
展开▼