Universal-Phonetic-Segment-Based Speech Coding and Its Applications to Speech Processing

在这个项目中，我们提出了一个新的语音处理框架，其中所有的声学语音样本都被编码成通用语音段（UPS）序列，并且语音文档处理（SDP）系统，如识别、检索、索引，都是在这个UPS域上构建的。采用这种框架，SDP系统与原始声学相关物或环境分离开来。这使得实现这样的灵活性成为可能，即识别型处理可以通过计算UPS序列之间的距离来处理，也可以构建在分布式处理方案上。通过本项目，我们在该框架上开发了以下组件技术：1)将原始的精细次语音段（SPS）集作为UPS集，带来了多语言语音的高性能识别和简单处理；2)有效的基于DP（动态规划）的序列匹配算法，称为Shift CDP和Relay CDP。通过构建语音识别和开放词汇口语文档检索（SDR）系统，评估了处理框架、SPS集和基于dp的算法的有效性。实验结果表明，该系统在性能评价方面优于传统方法。我们最终构建了一个实时开放词汇SDR系统进行演示，该系统可以通过用户的语音检索广播视频。

In this project, we present a novel speech processing framework, where all of the acoustic speech samples are once encoded into universal phonetic segment (UPS) sequences and spoken document processing (SDP) systems, such as recognition, retrieval, indexing, are constructed on this UPS domain. Adopting this framework, the SDP systems are separated from the original acoustic correlates or environments. This makes it possible to realize such flexibility that recognition-type processing can be handled by just calculating distances between UPS sequences, and also can be constructed on distributed processing schemes.Through this project, we have developed the following component techniques on this framework : 1)an original fine sub-phonetic segment (SPS) set as the UPS set, which brought high performance recognition and easy processing of multilingual speech, 2)effective DP(dynamic programming)-based sequence matching algorithms, called Shift CDP and Relay CDP. Effectiveness of the processing framework, the SPS set, and DP-based algorithms are evaluated by constructing speech recognition and open vocabulary spoken document retrieval (SDR) systems. Experimental results showed that the proposed SDP systems are superior to those based on conventional methods in performance evaluation. We have finally constructed a real time open vocabulary SDR system for demonstration, in which the system can retrieve broadcast video by user's speech.