LANDMARK ANALYSIS OF DYSARTHRIC SPEECH

点击翻译按钮获取中文摘要

The speech of individuals with dysarthria is highly variable and causes even trained, speaker-dependent speech recognizers to fail to recognize dysarthric speech. A novel method is proposed for recognizing words from a very small vocabulary of dysarthric speech. The method uses a "landmark"-based classifier (LMC) to identify sequences of acoustic events in speech. This new method does not depend on a stored database of speech. For small word sets, each landmark series will correspond to a unique word or phrase; the numbers from 1-20 and from 20-100 by tens form such a set. Small vocabularies are useful for applications such as speech interfaces for remote telephone and appliance control. Remote control systems are useful for motorically individuals and increase independence and safety. This project will extend the result for numbers spoken by normal subjects to more words and to some moderate dysarthrics. It will also determine whether these landmark series can be detected automatically in dysarthric speech by an LMC. If they can be detected robustly, Phase II will extend the analyses to a wider variety of dysarthric speech and a larger vocabulary, appropriate for remote control. PROPOSED COMMERCIAL APPLICATION: There is a clear need for speech recognition technology that will recognize disordered speech. The software could be used as a hands-free interface for AAC devices or environmental controllers.