A Model for Determining Consonantal Features in Speech

这个项目的长期目标是开发一个模拟人类对口语中单词的感知的模型。本项目的近期目标是完成这项任务的一部分：识别口语音节、单词和句子中辅音的独特特征。我们方法的关键思想是，即使表面语音是可变的，语音信号也包含说话人实现发音的可靠声学线索，并且发音由抽象的对比音位表示来管理。辅音特征的识别通过首先执行语音的声学分析以确定声音中形成辅音闭合和释放的标志或间断的位置来进行。然后，对这些地标附近的声音进行进一步的详细分析，以建立产生每个地标的辅音的基本特征，包括发音位置、发音特征、鼻音特征等。该进一步分析涉及从声音中提取为每个基本辅音特征提供线索的多个属性。这些属性的选择遵循这样的要求，即它们与产生语音的发音形状和运动密切相关。在拟议的工作中，我们目前对最有效地揭示发音及其支配特征和分段的属性组合的理解将通过详细的理论驱动的声学测量、感知实验和适当的统计分析来扩展和细化。模型的稳健性将使用从引用形式到连续演讲的各种话语进行评估。该模型的性能还将在已被噪声污染的语音中进行测试，并将该模型所产生的错误与人类听者所做的错误进行比较。该模型可用于研究听力受损的听者或在语音退化环境中的听者的言语知觉。对这些语音感知过程的理解可以导致改善语音感知和产生障碍的补救方法。

The long-range goal of this project is to develop a model that simulates the human perception of the words in spoken language. The immediate goal in the present project is to achieve one part of this task: the identification of the distinctive features of consonants in spoken syllables, words and sentences. Key ideas in our approach are that the speech signal contains reliable acoustic cues to the articulation implemented by the speaker even when surface phonetics are variable, and that the articulation is governed by abstract contrastive phonological representations. Identification of the consonant features proceeds by first performing an acoustic analysis of the speech to establish the locations of landmarks or discontinuities in the sound where consonant closures and releases are formed. The sound in the vicinity of those landmarks is then subjected to further detailed analysis to establish the underlying features of the consonant that generates each landmark, including the place of articulation, the voicing feature, the nasal feature etc. This further analysis involves extracting from the sound a number of attributes that provide cues for each of the underlying consonant features. The selection of these attributes is guided by the requirement that they be closely related to the articulatory shapes and movements that produced the speech. In the proposed work, our current understanding of the combination of attributes that most effectively reveal the articulation and its governing features and segments will be expanded and refined, through detailed theory-driven acoustic measures, perceptual experimentation and appropriate statistical analysis. The robustness of the model will be evaluated using various kinds of utterances, from citation forms to running speech. The performance of the model will also be tested in speech that has been contaminated with noise, and the errors made by the model will be compared with those made by human listeners. The model has application in the study of speech perception by listeners with impaired hearing or by listeners in an environment in which speech is degraded. Understanding of these processes of speech perception can lead to improved approaches to the remediation of disorders of speech perception and production.