Study on spoken language understanding framework integrating knowkedges among multiple layers

本研究的重点是开发一个框架，以统计的方式将从语音信号处理到口语理解的多个知识层的处理直接集成到语音识别过程中。语言模型、声学模型和对话模型三层的统计模型得到了广泛的研究。在集成方面，研究了基于贝叶斯风险最小化的语音解码方法，其中所有约束都可以表示为贝叶斯风险，以及利用语音信息进行对话管理和轮询的集成方法。部分结果作为开源语音交互构建工具MMDAgent和Julius的一部分公开提供。

This study focuses on developing a framework that integrates handling of multiple knowledge layer from speech signal processing to spoken language understanding directly into speech recognition process in a statistical mannar. Statistical models at layers of language model, acoustic model and dialogue model are widely investigated. For integration, speech decoding based on Bayes-risk minimization in which all the constraint can be expressed as Bayes risk, and some integration methods that utilizes speech information for dialogue management and turn taking was investigated. Part of the results are publicly available as part of an open-source voice interaction building tool MMDAgent and Julius.