Understanding Prosody and Tone Interactions through Documentation of Two Endangered Languages

为了实现文本到口语或口语到文本的准确自动翻译，有必要了解语音的韵律模式，因为韵律，即重音，节奏和短语之间的停顿，是传达意义和构建话语的核心。此外，许多语言使用声调来改变单词的意思，但韵律特征如何与声调相互作用却鲜为人知。由于音调和韵律是通过使用一些相同的声学维度（音高、持续时间、语音质量）来传达的，因此必须知道这两个系统如何相互作用以实现自动语音识别和翻译。为了调查和测试音调和韵律关系的假设，克里斯蒂安·迪·卡尼奥和一个跨学科团队将创建一个数据库，其中包含30小时的伊图约索Trique的转录叙述，并扩展Yoloxóchitl米斯特克语现有的类似数据库，这两种语言都来自墨西哥中东部的米斯特克语。然后，这些转录的叙述将被解析成更小的单元，应用自动语音识别中使用的“强制对齐”工具。项目的重要成果将包括对强制对准工具的测试和改进；和Yoloxóchitl Mixtec的新语料库和扩展字典；并分析了这两种语言的语调和韵律的相互作用。由此产生的韵律分段和标记语料库将是一种濒危语言的首个此类语料库，它将为讲话者社区和更广泛的科学界所用。该项目部分由稳健情报计划提供资金支持。

For accurate automated translation of text-to-spoken-word or spoken-word-to-text, it is necessary to understand the prosodic patterning of speech since prosody, i.e. stress, rhythm, and pauses between phrases, is central in conveying meaning and structuring discourse. In addition, many languages use tone to change the meaning of words but little is known about how prosodic features interact with tone. Since tone and prosody are conveyed by using some of the same acoustic dimensions (pitch, duration, voice quality), one must know how these two systems interact for automatic speech recognition and translation.To investigate and test hypotheses about the relationship of tone and prosody, Christian Di Canio, along with an interdisciplinary team, will create a database of 30 hours of transcribed narratives from Itunyoso Trique and expand an existing similar database in Yoloxóchitl Mixtec, both both Mixtecan languages from eastern central Mexico. These transcribed narratives will then be parsed into smaller units applying a 'forced alignment' tool used in automatic speech recognition. Important results from the project will include the testing and improvement of the forced alignment tool; new corpora and expanded dictionaries for and Yoloxóchitl Mixtec; and an analysis of tone and prosody interactions in these two languages. The resulting prosodically segmented and tagged corpora will be a first of its kind for an endangered language that will be of use to the speaker-community and the broader scientific community.This project is partially supported by funds from the Robust Intelligence program.