Doctoral Dissertation Research: The Association and Alignment of French Rises and their Role in Discourse Segmentation

在玛丽·E·贝克曼博士的指导下，波林·韦尔比将为她的博士论文收集法语语调上升的数据，以此作为识别一个单词的结束和另一个单词的开始的线索。这样的提示是必要的，因为演讲缺乏文字通常具有的单词之间的空间。在英语中，重音是一个很好的切分线索，因为实词(名词、形容词等)通常以重读音节开头。法语没有这样的重音提示，但人们假设法语中语调上升的时机也是为了填补类似的角色。以前的研究表明，法语语调的特点是在短语结束时必须上升，在接近开始时可以选择提前上升。两个生产实验将考察对早期上升的影响(例如，语速)以及两次上升的时间。三个感知实验将检验早期上升在语音切分中的作用。这些项目将由以英语为母语的人阅读并记录下来；然后使用计算机重新合成来改变上涨的时间。参与者将倾听刺激并做出各种类型的反应。例如，一个实验将使用虚构的非单词，其切分反映了未知但真实的单词的切分。例如，相同的音节序列可以解释为无意义的单词“Melamondine”或短语“me lamondines”(‘my’加上无意义的单词“lamondine”)。根据一种假说，从第一个音节开始的早起会使听者偏向于一个单词的解释(“Melamondine”)，而从第二个音节开始的早起会使听者偏向于两个单词的解释(“me lamondines”)。这一研究对理论语言学和语音技术的发展都具有重要意义。它将增加我们对法语语调和语音切分的理解。此外，更好地理解分割策略可以改进自动语音识别系统，后者在分割准确率方面远远落后于人类。更精确地描述计时事实也可能改善语音合成系统，因为糟糕的语调会使合成的语音难以理解。反过来，语音技术的进步为科学家未来的语音研究提供了更好的工具。

Under the direction of Dr. Mary E. Beckman, Ms. Pauline Welby will collect data for her doctoral dissertation on intonational rises in French as cues to identifying where one word ends and another begins. Such cues are necessary since speech lacks the spaces between words that writing often has. In English, stress is a good cue to segmentation since content words (nouns, adjectives, etc.) typically begin with stressed syllables. French lacks such stress cues, but the timing of intonational rises there has been hypothesized to fill a similar role. Previous research has shown that French intonation is characterized by an obligatory rise at the end of a phrase and an optional early rise near the beginning. Two production experiments will examine influences on the early rise (e.g., speaking rate) and the timing of both rises. Three perception experiments will examine the role of the early rise in speech segmentation. Items will be read by a native speaker and recorded; the timing of the rises will then be varied using computer resynthesis. Participants will listen to stimuli and make various types of responses. For example, one experiment will use invented non-words, whose segmentation reflects that of unknown but real words. For example, the same sequence of syllables could be interpreted as the nonsense word "melamondine" or as the phrase "mes lamondines" ('my' plus the nonsense word "lamondine"). According to one hypothesis, an early rise beginning at the first syllable will bias the listener to the one-word interpretation ("melamondine"), while an early rise beginning at the second syllable will bias the listener to the two-word interpretation ("mes lamondines"). This research is significant both for theoretical linguistics and for the development of speech technologies. It will increase our understanding of French intonation and of speech segmentation. In addition, a better understanding of segmentation strategies can improve automatic speech recognition systems, which lag far behind humans in segmentation accuracy. A more refined account of the timing facts might also improve speech synthesis systems, since poor intonation makes synthesized speech difficult to understand. In turn, improvements in speech technology provide scientists with better tools for future speech research.