Automatic indexing for lecture speech and its advanced utilization through speech interaction

我们收集了16名发言者、114场讲座、3860分钟的课堂演讲，并发布了语料库。我们开发了我校研究生课堂讲稿数据的自动语音识别、句子提取、切分/标引、语音检索和讲课浏览系统的构建过程。这些过程是必要的，以提高播放声音或视频数据的可用性，在讲课、总结和索引讲稿或视频的情况下，使学生能够更有效地学习。我们的目标是构建这样一个有条理的授课内容框架。为了实现这一目标，我们首先研究了录音方法对语音识别性能的影响。事实证明，高质量的手持麦克风和低质量的翻领麦克风在准确度上存在23%的差异。此外，我们利用相关的Web文本对领域相关的语言模型进行了改进，并提出了一个填充词插入模型。其次，我们尝试通过提取重要句子来进行自动摘要，我们得到了0.319-0.456的κ值，与人类的0.407-0.477不相上下。最后，我们构建了一个讲座浏览系统，使用户能够更有效地利用上述过程的结果进行学习，并对其进行了评估

We collected the class room lecture speech consisting of 16 speakers, 114 lectures, and 3860 minutes, and publised the corpus. We developed the procedure of automatic speech recognition, sentence extraction, segmentation/indexing, spoken retrieval and construction of lecture browsing system for classroom lecture data of our university's graduated course. These processes axe necessary to improve the usability of broadcasting sound or video data In the case of lecture, summarized and indexed lecture speech or video enables to students to more effective leaning. Our goal was to construct a framework of such structured lecture contents. To achieve this goal, first, we investigated influence of the recording methods on the speech recognition performance. It turned out that there was 23% difference on the accuracy between a high quality hand-microphone and a low quality lapel microphone. Furthermore, we improved the domain-dependent language model by using related Web texts and developed a filler insertion model. Second, we tried automatic summarization by extracting important sentences, and we obtained 0.319-0.456 κ value, comparable with human doing 0.407-0.477. Finally, we constructed the lecture browsing system which enables users to learn more effectively by using results of the procedure described above, and evaluated it