Prosodic Event Annotation and Detection in Three Varieties of English

说英语的人经常毫不费力地使用韵律，比如口音和措辞，来进行超越单个单词含义的交流。然而，我们对韵律的理解，以及它与语音技术的结合，已经落后了，特别是对于那些已经被低估的语言品种。虽然数以百万计的美国人说非洲裔美国英语（AAE）或拉丁裔英语（LE），但这些语言在基础和应用研究中的代表性仍然不足，对语言治疗、教育、刑事司法和就业产生了负面影响。本研究将产生一个由AAE， LE和欧美英语语音数据标注的韵律语料库，用于提高我们对韵律的科学理解，并创建自动检测重音和短语边界的工具。一个由来自不同背景的7名学生组成的团队将接受培训，这将为语言和技术职业提供一个门户。语料库将包括在地图任务中记录的同一方言的对话人，这是一种用于引出自发的、自然的语言的方法。数据将使用自动语音识别和手动校正进行转录，并使用强制校准软件自动分割成单词和单个声音。训练有素和未经训练的编码员都将使用快速韵律转录来注释句子重音和边界数据，这是一种为非专家快速，直观的韵律注释而设计的方法。声学信息，包括持续时间、音高和强度，将被提取出来，并与注释一起用于训练机器学习检测模型，并测试之前关于三种英语语言变体的频率、分布、声学和口音和边界感知的发现。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

English speakers constantly and effortlessly use prosody, such as accent and phrasing, to communicate beyond the meaning of individual words. Yet our understanding of prosody, and its integration in speech technologies, have lagged behind, particularly so for language varieties that are already underrepresented. While millions of Americans speak African American English (AAE) or Latinx English (LE), these language varieties remain underrepresented in basic and applied research, with negative consequences for speech therapy, education, criminal justice, and employment. This research will result in a corpus of AAE, LE and European American English speech data annotated for prosody, which will be used to improve our scientific understanding of prosody and to create tools for automatic detection of accent and phrase boundaries. A team of 7 students from diverse backgrounds will receive training that will provide a gateway to careers in language and technology.The corpus will include pairs of same-dialect speakers recorded in a map task, a method used to elicit spontaneous, naturalistic speech. The data will be transcribed using automatic speech recognition and hand correction, and automatically segmented into words and individual sounds using forced alignment software. Both trained and untrained coders will annotate the data for sentence accent and boundary using Rapid Prosody Transcription, an established method designed for fast, intuitive prosodic annotation by non-experts. Acoustic information, including duration, pitch, and intensity, will be extracted and used with the annotations to train a machine learning detection model, and to test previous findings about the frequency, distribution, acoustics and perception of accent and boundary in the three English language varieties.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.