Study on Multimedia Teaching Material by Hyper-Media and Content Analysis of Lecture Videos

本研究从1999年到2001年的主要研究成果可以归纳为以下7点：1。基于高速、高精度语音识别的语音文档索引：提出了基于音素或单词错误最小化的语音解码以及无监督模式下声学模型的迭代自适应。基于噪声和BGM鲁棒语音识别的噪声语音索引：利用卡尔曼滤波和mlr .3提出了基于非平稳和平稳降噪的语音识别方法。基于说话人识别的说话人索引：提出了基于音素和说话人分离的说话人识别方法，并在此基础上对个人进行索引。基于字符和图像识别的视频图像索引：通过视频字幕帧选择、有效二值化和OCR.5，提出视频字幕或翻转识别。基于语音和字符识别的话题分割：基于本文提出的词空间方法，将新闻视频分割为单个话题。在商业视频中，将视频字幕识别与语音识别相结合，提出了主题分割。7.视频图像的构建与内容描述：通过索引和主题分割生成视频图像的目录。针对超链接构建的主题检索和总结：提出了跨媒体检索，提出了特定人物针对特定主题发言的视频片段提取。

Main results of this study from 1999 to 2001 are summarized into seven points as follows;1. Indexing to spoken documents based on high speed and high accurate speech recognition: Speech decoding based on phonemes or words error minimization was proposed as well as iterative adaptation of acoustic models in unsupervised mode.2. Indexing to noisy speech based on noise and BGM robust speech recognition: Speech recognition based on non-stationary as well as stationary noise reduction was proposed by using Kalman filter and MLLR.3. Speaker indexing based on speaker recognition: Speaker recognition based on phoneme and speaker separation was proposed and individual person was indexed based on the proposed method.4. Indexing to video image based on character and image recognition: Video caption or flip recognition was proposed by carrying out the video caption frame selection, effective binarization and OCR.5. Topic segmentation based on speech and character recognition: News videos were segmented into individual topic based on word space method proposed in this study. In commercial video, the topic segmentation was proposed by integrating video caption recognition and speech recognition.6. Structuring and content description to video image: Table of contents of video images was produced after indexing and topic segmentation.7. Topic retrieval and summarization for hyperlink construction: Cross media retrieval was proposed as well as the video clip extraction where the specific person was speaking about the specific topics.