MRI: A Feature - Based Word Recognition System

该项目的目标是开发一种方法来实现一个特征--基于词的识别系统，可以应对词的声学表现中可能出现的广泛的变化性。词汇声学模式的差异有许多不同的来源，包括说话人内部和说话人之间在说话风格和语速上的差异，各种语境效果，以及说话人之间的解剖差异。因此，虽然语言描述使用二进制值特征，但特征的声学相关性具有随时间变化的强度和程度。研究将分为三个阶段。为确保充分反映可能发生的变化，将设计一个约200字的数据库。测试词将从不同的语境中分离出来，由几个说话者说出，无论是男性还是女性。接下来，将对语料库进行声学研究，以补充文献中关于特征的声学相关性的数据，并进一步研究可变性。最后，将开发一个不依赖于说话人的孤立词识别系统，并使用一个新的200个单词的数据库进行测试。识别系统将由三个组件组成：用于提取特征的声学属性的属性检测器、将由每个单词的特征的某种排列组成的词典、以及用于在提取的声学模式和词汇表示之间进行映射的匹配器。将有几种方法来处理可变性。首先，特征的声学属性将基于相对测量，因此它们将最大限度地依赖于说话人的特征。第二，不同于传统的基于特征的识别系统，词汇访问将直接从声学属性而不是假设的语音片段来实现。这样，在匹配过程中可以更直接地考虑上下文信息。最后，研究人员将开发一个框架来表达单词的词汇表征中的可变性。尽管可变性有几个来源，但由此产生的声学变化似乎可以通过几个特征改变过程来表达，包括语音同化、特征同化和其他一些结构相关的变体。因此，通过将特征改变过程结合到单词的词汇表示中，识别系统在能够识别特定单词之前，不需要对可能出现在特定单词中的所有可变性进行“看到”或具有规则。这个项目的结果应该有助于我们对声学语音学的理解，也应该有助于改进自动语音识别技术的发展。

ABSTRACT The aim of the project is to develop an approach to a feature- based word recognition system which can cope with the wide range of variability that can occur in the acoustic manifestation of words. The variability in the acoustic pattern of words has many different sources, including intra- and inter-speaker differences in speaking style and rate, various contextual effects, and anatomic differences between speakers. Thus, whereas linguistic descriptions use binary-valued features, the acoustic correlates of the feature have varying degrees of strength and extent over time. The research will be divided into three stages. To ensure an adequate representation of the variability that can occur, a database of around 200 words will be designed. The test words will be excised from a variety of contexts and spoken by several talkers, both males and females. Next, an acoustic study of the corpus will be conducted to supplement data in the literature on the acoustic correlates for features and to further study variability. Finally, a speaker-independent isolated word recognition system will be developed and tested with a new database of the same 200 words. The recognition system will consist of three components: property detectors for extracting the acoustic properties for features, a lexicon which will consist of some arrangement of features for each word, and a matcher for mapping between the extracted acoustic patterns and the lexical representation. There will be several ways to handle variability. First, the acoustic properties for features will be based on relative measures so that they will be minimally dependent upon the speaker's characteristics. Second, unlike traditional feature- based recognition systems, lexical access will be achieved directly from the acoustic properties rather than from postulated phonetic segments. In this way, contextual information can be taken into account more directly during the matching process. Finally, the investigator will develop a framework for expressing variability within the lexical representation of words. Although variability has several sources, it appears that the resulting acoustic changes can be expressed in terms of a few feature- altering processes including phonetic lenition, feature assimilation and some other structurally-dependent variants. Thus by incorporating the feature-altering processes in the lexical representation of words, it should not be necessary for the recognition system to have "seen" or have a rule for all of the variability that can occur in a particular word before it is able to recognize it. The results of this project should contribute to our understanding of acoustic phonetics, and also to the development of improved technologies for automatic speech recognition.