VERFAHREN ZUR SPRECHERADAPTIVEN ERKENNUNG VON SPRACHE

摘要

A method of this type is intended to be suitable for the recognition of individual words as well as continuous speech. It is to be characterised by robust word pattern recognition in the face of incorrect syllable segmentation and variable diction, e.g. when syllables are slurred. It must also enable fast adaptation of the system to a new speaker and an essentially unlimited generation and extension of the vocabulary from written text without explicit system training through pronunciation. An approximately real-time recognition of words and word sequences should also be possible even with very extensive vocabularies. Known methods of speech recognition require a highly expensive training procedure. Furthermore, an unmanageable flow of hypotheses is generated with continuous speech and a relatively large vocabulary, and even with medium-sized vocabularies. �According to the invention, the extracted feature vectors are segmented and classified into a set of hypotheses comprising syllable-oriented word sub-units with the aid of a word sub-unit inventory. A word hypothesis set, which is subjected to syntactic analysis in order to identify the spoken phrase, is generated from the hypothesis set by means of a three-dimensional, time-dynamic comparison with diction variants from a reference pattern vocabulary. ��