Efficient configuration of robust spoken language technology based human interfaces

自动语音识别(ASR)算法和实现的发展导致了围绕该技术构建的大量应用程序和服务。ASR技术出现在已经开发用于人机交互、语音口述、基于语音的音频和视频资源索引的系统以及用于提高各种其他人类语言相关活动的效率的计算机辅助系统中。然而，与当今存在的技术相关的两大限制。首先，众所周知，说话人群体、说话风格、声学环境和通信渠道的变化是脆弱的。其次，众所周知，这项技术也很昂贵，因为成本主要是为新应用程序和新语言开发系统所需的大量转录语音数据。这个项目从几个方面解决了这些缺点。首先，将开发有效的模型表示和训练算法，以使为新任务和新语言配置ASR系统的过程成本更低。它们将促进这样的场景，即潜在的巨大但廉价的与应用无关的语音语料库可以与小得多、更昂贵的特定任务语音语料库一起使用来配置ASR系统。设想了多语言场景，其中可以使用来自多种语言的语音数据来训练针对几乎不存在语音或文本语料库的语言的ASR系统。其次，将开发基于人类语音产生和语音感知模型的说话人归一化和声背景补偿方法。这将导致更有效的范例，以消除以前看不见的扬声器或声学环境的影响。最后，这些技术将被应用于创建更健壮的基于口语技术的人机界面。主要目标是开发和评估使能技术，主要动机是将基于文本的应用程序(例如，在线搜索引擎)的能力扩展到口语界面。

Developments in automatic speech recognition (ASR) algorithms and implementations have resulted in a remarkable number of applications and services built around this technology. ASR technology appears in systems that have been developed for human-machine interaction, speech dictation, speech based indexing of audio and video resources, and computer aided systems for increasing the efficiency of various other human language related activities. However, there are two major limitations associated with the technology as it exists today. First, it is known to be fragile with respect to variation in speaker populations, speaking styles, acoustic environments, and communications channels. Second, the technology is also known to be expensive where the cost is dominated by the large quantities of transcribed speech data needed for developing systems for new applications and new languages. This project addresses these short-comings from several directions. First, efficient model representations and training algorithms will be developed to make the process of configuring ASR systems for new tasks and new language less expensive. They will facilitate scenarios where potentially huge but inexpensive application independent speech corpora can be used with much smaller more expensive task specific speech corpora to configure ASR systems. Multilingual scenarios are envisioned where speech data from many languages can be used for training an ASR system for languages where little speech or text corpora exist. Second, speaker normalization and acoustic background compensation approaches motivated by models of human speech production and speech perception will be developed. This will lead to more efficient paradigms for eliminating the effects of previously unseen speaker or acoustic environments. Finally, these techniques will be applied to creating more robust spoken language technology based human interfaces. The major goal is to develop and evaluate enabling technology with a major motivation being to extend the capabilities of text based applications like, for example, online search engines, to spoken language interfaces.