RI: Medium: Collaborative Research: Explicit Articulatory Models of Spoken Language, with Application to Automatic Speech Recognition

该奖项是根据2009年美国复苏和再投资法案（公法111-5）资助的。自动语音识别的主要挑战之一是说话风格的可变性，包括说话速度的变化和协同发音。发音器官（如嘴唇和舌头）的模型可以简洁地表示这种变化。先前关于发音模型的大部分工作都集中在声学和发音之间的关系上，但更重要的改进需要隐藏发音状态结构的模型。这项工作既有提高识别的技术目标，也有更好地理解发音现象的科学目标。该项目考虑了比以前研究过的更大的模型类。特别是，该项目开发了图形模型，包括动态贝叶斯网络和条件随机场，旨在利用发音知识。认识到有向和无向混合图形模型以及生成和判别训练的好处，目前正在开发一个有向和无向混合图形模型的新框架。项目活动包括使用上下文建模、异步结构和专门培训对早期铰接模型进行主要扩展；铰接变量因子条件随机场模型的建立以及辨别性训练，以减轻词语混淆。科学目标解决的问题，其中发音轨迹在不同的情况下变化的方式。使用了现有的数据库，并扩展了手工发音注释方面的初步工作。此外，该项目使用发音模型来执行大型数据集的强制转录，为研究界提供了额外的资源。其他广泛的影响包括适用于其他时间序列建模问题的新模型和技术。扩大语音识别的适用性将有助于它实现更有效地存储和获取语音信息的承诺，并为听力或运动障碍者提供平等的技术竞争环境。

This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).One of the main challenges in automatic speech recognition is variability in speaking style, including speaking rate changes and coarticulation. Models of the articulators (such as the lips and tongue) can succinctly represent much of this variability. Most previous work on articulatory models has focused on the relationship between acoustics and articulation, but more significant improvements require models of the hidden articulatory state structure. This work has both a technological goal of improving recognition and a scientific goal of better understanding articulatory phenomena.The project considers larger model classes than previously studied. In particular, the project develops graphical models, including dynamic Bayesian networks and conditional random fields, designed to take advantage of articulatory knowledge. A new framework for hybrid directed and undirected graphical models is being developed, in recognition of the benefits of both directed and undirected models, and of both generative and discriminative training. The project activities include major extension of earlier articulatory models with context modeling, asynchrony structures, and specialized training; development of factored conditional random field models of articulatory variables; and discriminative training to alleviate word confusability.The scientific goal addresses questions about the ways in which articulatory trajectories vary in different contexts. Existing databases are used, and initial work in manual articulatory annotation is being extended. In addition, the project uses articulatory models to perform forced transcription of larger data sets, providing an additional resource for the research community. Other broad impacts include new models and techniques with applicability to other time-series modeling problems. Extending the applicability of speech recognition will help it fulfill its promise of enabling more efficient storage of and access to spoken information, and equalizing the technological playing field for those with hearing or motor disabilities.