Collaborative Research: Effects of production variability on the acoustic consequences of coordinated articulatory gestures

人类的声音是我们最古老、最可靠的交流工具，现在正迅速成为我们每天与汽车导航系统、医疗和法律听写系统、像Siri这样的个人助理、自动金融系统等技术进行交互的首选输入界面。成千上万的“应用程序”已经被开发出来，以帮助消费者使用语音来获取他们正在寻找的信息。语音识别是所有这些技术的支柱。因此，语音识别器的性能是客户满意度的关键。目前，许多系统仍然需要对特定的说话者进行调整才能表现良好，并且识别任务必须在其他方面受到限制，例如要求(1)使用特定的词汇，(2)大多数单词，特别是内容单词的发音清晰，以及(3)限制背景噪声。在本研究中，将研究语音变异性，并开发方法和模型，使识别器更加独立于说话者，能够处理从清晰发音到非常随意的语音的各种语音风格。研究结果还将涉及言语计划和组织的语言模型，为说话者如何在言语生产效率与可理解性之间进行权衡提供证据。在这个项目中，语音发音器的点源跟踪将与相应的声学同时收集。讲话者将以正常和快速的节奏记录语音（后者的目的是显著增加信号的变化程度）。这些数据将允许调查说话者是否总是在期望目标的方向上移动他们的发音器（例如，在发/t/时舌尖到牙齿），即使快速的生产速度遮挡了相关的声学信息（如“完美”）。如果得到证实，这一发现将为通过结合发音信息使识别系统更加强大指明道路。此外，这些数据将支持语音反转系统的开发，该系统能够“发现”隐藏的发音运动，这些运动可能被声学掩盖。

The human voice, our oldest and most reliable communication tool, is now rapidly becoming the input interface of choice that we use everyday to interact with technologies such as car navigation systems, medical and legal dictation systems, personal assistants like "Siri," automated financial systems, etc. Thousands of 'apps' have been developed to help consumers use voice to get the information they are looking for. Speech recognition is the backbone of all of these technologies. As a result, the performance of speech recognizers is key for customer satisfaction. Currently, many systems still need to be tuned for a particular speaker to perform well, and the recognition task has to be limited in other ways such as requiring (1) usage of a specific vocabulary, (2) clear pronunciation of most of the words, especially the content words and (3) limited background noise. In this research, speech variability will be studied, and methods and models will be developed that will enable recognizers to be more speaker independent and capable of handling the full range of speech styles from clear articulation to very casually spoken speech. The results will also bear on linguistic models of speech planning and organization, providing evidence for how speakers trade off efficiencies in the production of speech against the need to be intelligible.In this project, point source tracking of the speech articulators will be collected concurrently with the corresponding acoustics. Speakers will record speech at both a normal and rapid pace (the purpose of the latter is to increase significantly the degree of variability in the signal). This data will allow for the investigation of whether speakers always move their speech articulators in the direction of a desired target (e.g. tongue tip to teeth in producing /t/) even when a rapid production pace occludes the relevant acoustic information (as in "perfect"). If confirmed, this finding will point the way towards making recognition systems more robust through the incorporation of articulatory information. In addition, such data will support the development of a speech inversion system capable of 'uncovering' hidden articulatory movements potentially masked from the acoustics.