CISE Research Instrumentation: A Study of an Event Based and Feature Based Approach to Speech Recognition

96-17426 Espy-Wilson，Carol Y. 波士顿大学 CISE研究工具：基于事件和基于特征的语音识别方法研究 本研究的目的是开发一个基于语音特征的语音识别系统，并利用该系统的开发作为理解变异性的工具。 相对于倒谱参数，宽类识别结果表明，基于语音特征的信号表示能够更好地针对语言信息，并且它以与说话者无关的方式这样做。 此外，这样的信号表示允许语音过程的直接研究。 拟议的设备将有助于扩大这项工作。首先，在计算时间的急剧削减所提供的超2-将加快开发和优化的声学参数使用费舍尔准则和自动分类器，-使可行的实验与一个完整的面向对象的词典，-允许实验与计算昂贵的算法提取的声学事件和-允许更快的计算识别结果。 第二，额外的磁盘将允许调查更大的数据集和在线存储计算参数。 第三，工作站将通过允许用户收听话语并在本地执行处理来帮助分析。 最后，拟议的PC和合成器将允许通过声学建模和分析合成的可变性的研究。

96-17426 Espy-Wilson, Carol Y. Boston University CISE Research Instrumentation: A Study of an Event Based and Feature Based Approach to Speech Recognition The aims of this study are to develop a speech recognition system based on phonetic features and to use the development of this system as a tool for understanding variability. Relative to cepstral parameters, broad-class recognition results show that a signal representation based on phonetic features is better able to target the linguistic information and it does so in a more speaker-independent manner. Furthermore, such a signal representation allows for a direct study of phonological processes. The proposed equipment will help extend this work. First, the drastic cut in computation time afforded by the Ultra 2- will speed up the development and optimization of acoustic parameters using the Fisher criterion and an automatic classifier, - make feasible experimentation with a full object-oriented lexicon,- allow experimentation with computationally expensive algorithms for the extraction of acoustic events and -allow faster computation of recognition results. Second, the extra disks will allow investigation of larger data sets and the storage of computed parameters on line. Third, the workstations will aid analysis by allowing users to listen to utterances and perform processing locally. Finally, the proposed PC and synthesizer will allow a study of variability through acoustic modeling and analysis-by-synthesis.