Three dimensions of sentence prosody

句子韵律的三个维度

• 批准号: RGPIN-2018-06153
• 负责人: Wagner, Michael
• 金额: $ 5.83万
• 依托单位: McGill University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=758104
• 关键词: Three; dimensions; sentence; prosody

Speech synthesis has come a long way. Today's computers can produce words and sentences highly accurately, but if a computer reads out a novel text, the result is still often hard to listen to. A major reason is our lack of understanding of sentence prosody. Sentence prosody can be characterized as all those acoustic properties of an utterance that are not a function of the words it contains, but rather are due to other factors, and includes three separate dimensions: Intonation (e.g., falling intonation on assertions vs. rising intonation on questions), Phrasing (the way the words in a sentence group together), and Prominence (for example highly predictable or discourse given words are often reduced, while contrastive information is boosted). These three dimensions of sentence prosody (intonation, phrasing, prominence) are partly determined by the phonological representations of a particular language (e.g., declarative intonation in French sounds different than in English) and partly a direct reflex of on-line processing during speaking (e.g., complex constituents are often preceded by phrasing junctures). Once we understand the representational questions and can disentangle the three dimensions, sentence prosody can be used as a window into cognitive mechanisms underlying speech production. This proposal brings together several strands of research that my lab has contributed to in the past.The first step to a better understanding of sentence prosody is to disentangle the three mentioned dimensions in the signal. The syntactic and semantic functions (type of speech act, grouping, contrast) are orthogonal to each other, but whether their prosodic correlates (tune, phrasing, prominence) are remains controversial. Most current models of sentence prosody assume interactions or even conflate the dimensions. However, the reality of the assumed interactions between the dimensions has not been systematically established. An initial study, conducted for this proposal in my lab (Wagner & McAuliffe 2017, Interspeech), systematically crossed the three functional dimensions (speech act, syntax, contrast) in a factorial speech production experiment. The results show that we can in fact recover the three `dimensions' of prosody quite reliably from speakers' utterances, once we factor out the contribution of the other dimensions, as well as speaker differences. While we find limited interactions, they are not actually the ones that current models predict. This project consists of a series of production and perception experiments designed to establish the true interactions, and develops a more appropriate representational model. We make use of novel tools (e.g, The Montreal Forced Aligner, developed in my lab, Mcauliffe et al. 2017, Interspeech) and statistical modelling, to develop models that can disentangle the dimensions, just as humans do when processing speech.

语音合成已经取得了长足的进步。今天的计算机可以高度准确地产生单词和句子，但如果计算机读出新颖的文本，结果通常仍然很难听。一个主要原因是我们对句子韵律缺乏了解。 句子韵律可以被描述为话语的所有声学特性，这些特性不是其所包含单词的函数，而是由其他因素引起，并且包括三个独立的维度：语调（例如，断言时的降调与问题时的升调）、短语（句子中的单词组合在一起的方式）和突出性（例如，高度可预测的或给定单词的话语通常会被减少，而对比信息则被减少）。 增强）。句子韵律的这三个维度（语调、短语、突出音）部分是由特定语言的语音表征决定的（例如，法语中的陈述语调听起来与英语不同），部分是由说话过程中在线处理的直接反射决定的（例如，复杂的成分通常前面有短语连接点）。一旦我们理解了表征问题并能够理清三个维度，句子韵律就可以作为了解语音生成背后的认知机制的窗口。该提案汇集了我的实验室过去做出的几项研究。更好地理解句子韵律的第一步是解开信号中提到的三个维度。句法和语义功能（言语行为类型、分组、对比）彼此正交，但它们的韵律相关性（曲调、措辞、突出程度）是否存在争议。当前大多数句子韵律模型都假设相互作用，甚至将维度混为一谈。然而，维度之间假设的相互作用的现实尚未系统地确定。我的实验室针对该提案进行了一项初步研究（Wagner & McAuliffe 2017，Interspeech），在阶乘语音生成实验中系统地跨越了三个功能维度（言语行为、语法、对比）。结果表明，一旦我们排除其他维度的贡献以及说话者差异，我们实际上可以相当可靠地从说话者的话语中恢复韵律的三个“维度”。虽然我们发现交互作用有限，但它们实际上并不是当前模型预测的交互作用。该项目由一系列旨在建立真实交互的生产和感知实验组成，并开发出更合适的表征模型。我们利用新颖的工具（例如，我的实验室开发的蒙特利尔强制对齐器，Mcauliffe 等人，2017 年，Interspeech）和统计模型来开发可以解开维度的模型，就像人类在处理语音时所做的那样。