A Study of Variability & A Feature-Based Approach to Speech Recognition

9310518 Espy-Wilson今年是为期三年的连续奖项的第一年，该奖项旨在调查说话人的可变性和基于特征的语音识别方法。说话人可变性研究从目前的知识到考虑自发的语音数据库，目的是为了识别目的进行分类和建模。预计这项研究将显示系统的可变性，并有可能在单词的词汇表示中表示它，从而减轻对识别系统训练的要求。词汇项和其可能的声学实现之间的更一般可能的一致性的现象，例如透镜现象的表示提供了一个例子。对关系词中的语音特征进行研究，以寻找可能降低可变性的来源，并揭示可能对其相关性起中心作用的剩余因素(希望较少)，以便促进基于特征的识别。

9310518 Espy-Wilson This is the first year of a three-year continuing award to investigate speaker variability and a feature-based approach to speech recognition. The speaker variability study builds from current knowledge to consideration of a spontaneous speech database with the purpose of categorization and modelling for recognition purposes. It is anticipated that systematic variability will show in this study and that it may be possible to represent it within the lexical representation of words, easing requirements for training of recognition systems. The representation of phenomena such as lenition provides an example of the more general possible agreement between a lexical item and its possible acoustic realizations. A study of speech features in relational terms is carried out to search for sources that may decrease variability and to disclose the remaining, hopefully fewer, factors that may be central to their correlates in order to facilitate feature-based recognition.