Acoustic Cues to Discourse Structure for Interactive Speech Skimming

该项目自动检测数字记录的讲座中的话语结构，并使用此结构来促进录音的快速浏览。 语篇结构是话题之间的关系，包括副话题和嵌入的话语。 表示结构（而不是内容）的片段的边界由停顿、音高范围变化和提示短语来传达。 语篇结构具有层次性，主、次话题有多个层次。 通过对讲座语料库的分析，可以定量地给出声学特征与语篇结构之间的相关性，为自动分词算法提供依据。 一个统一的方法合并这些多个声学线索的提示，以映射层次话语结构的时间偏移记录。 派生语篇结构用于浏览录音，在不同层次的细节和总结。 演示是高度交互式的，允许用户通过一个物理界面来操纵回放，该界面是迭代设计和评估的产物。 手势控制回放过程中的细节级别、速度和随机访问。 更有效地访问音频记录使得有可能管理来自更广泛来源的信息，并且在比传统计算机图形显示更多的情况下。

This project automatically detects discourse structure in digitally recorded lectures and uses this structure to facilitate rapid browsing of the recording. Discourse structure is the relationship between topics, including sub-topics and embedded remarks. Boundaries of the segments representing the structure (but not content) are conveyed by pauses, pitch range variation, and cue phrases. Discourse structure is hierarchical, with multiple levels of major and minor topics. Analysis of a corpus of lectures provides quantitative correlation between acoustic features and discourse structure, and the basis of an automatic segmentation algorithm. A unified approach merges hints from these multiple acoustic cues to map hierarchical discourse structure to time offsets in the recording. Derived discourse structure is used to browse the recording, at various levels of detail and summarization. Presentation is highly interactive, allowing the user to manipulate playback via a physical interface which is the product of iterative design and evaluation. Hand gestures control the level of detail, speed, an d random access during playback. More productive access to audio recordings makes it possible to manage information from a wider variety of sources and in more situations than conventional computer graphic displays.