Acoustic/Linguistic Aspects of Automatic Speech Recognition.

摘要

As word-level automatic speech recognition has improved, there has been an increased awareness of the need for more general procedures to recognize extended continuous speech utterances. This project has therefore addressed several problems related to the recognition of continuous speech. This has involved the development of a flexible interactive software system for acquiring, labeling, and analyzing continuous speech. This systems has been used to complete most of the other studies reported here, including (1) a statistical procedure for the automatic extraction of reliable formant measures from large-scale data bases, (2) an algorithm for segmenting continuous speech on the basis of spectral change, (3) a study of vowel formant distributions in relation to the inter-speaker normalization problem, (4)various studies of the properties of vowels and consonants, and (5) studies of the prosodic patterns of speech, particularly as they can guide the recognition of extended utterances. (Author)