TarDiS: Targets and Dynamics in Speech

点击翻译按钮获取中文摘要

Speech is a physical phenomenon: it can be represented as a series of vocal tract movements, or as sound waves. However, in order for speech to serve a communicative function, it also must include more abstract components, building blocks that pair these physical phenomena with linguistic meaning. For instance, the vowel /ai/ in a word like 'kite', can be represented in multiple ways. Physiologically, it involves tongue movement from a lower and more far back position to the high front corner of the mouth, but at the same time /ai/ is an abstract category, a family of sounds brought together by their joint meaningful function. The vowel /ai/ is a contrastive unit in English, one that makes words like 'kite' different from words like 'kit'. Most native speakers of English never need to ponder the nature of the category /ai/, yet they somehow acquire a systematic mapping that links the physical form of /ai/ with its meaning. To linguists, understanding this process holds the key to understanding how language works. It is by this type of mapping that a movement of the tongue acquires linguistic meaning. This research project sets out to probe the nature of vowel categories like /ai/, and more specifically what their abstract representations must contain in order to successfully model variation observed in speech. This is an old problem in speech science, which we intend to approach using a novel type of data: combination of computational modelling, measurements of tongue movement based on tracking such movement in a magnetic field, and images of tongue acquired with ultrasound, as well as the audio signal. We will record such data from twenty native speakers from a single speech community (West Yorkshire) pronouncing all vowels in their variety of English. We will then consider various abstract features that we can derive from the physical signal, and the minimum number of such features needed to successfully capture all vowel categories. For instance, we want to know whether vowels such as /ai/ are best modelled as consisting of one or two elements, and how much information about movement we need to encode in order to successfully model /ai/ within a vowel system.We will extend the implications of our findings to practical applications. Although vowel categories and vowel targets do not feature overtly in natural acquisition of language, they become important in situations where explicit speech instruction plays a role, such as speech therapy. The more we understand about the nature of speech sounds, the more successful we can be in teaching them. Therefore, we will present our findings on speech goals from a theoretical modelling perspective to clinical speech researchers and clinical practitioners, and engage them in a discussion of how such abstract speech targets can be used as tools in speech therapy.