Lecture speech summarization based on the key sentence extraction using prosodic changes

为了研究基于重音短语聚类的随机语音分量模型的有效性，利用观察到的语音分量模式和模型对口语重音短语的口音类型进行估计。引入了关于重音规则的知识，以减少潜在的重音类型变化。该模型对152个口语讲座进行了训练，并对CSJ （Corpus of Spontaneous Japanese）中的另外15个口语讲座进行了评估。结果表明，引入重音规则提高了重音类型估计的性能，有效地构建了随机FO模型。由于语音数据不适合快速扫描，因此期望开发演讲语音自动摘要。为了实现自动摘要，从语音数据中手工提取重要的句子或单词。17名被试被要求总结来自CSJ的20个语音数据。通过对抽取关键句的一致性分析以及抽取关键句与语音方差的相关性分析，证实了对于韵律变化较大的口语讲座，被试对抽取关键句的认同程度较高。分析了200毫秒以上停顿自动分割的话语单元的重要性与韵律参数之间的关系。话语单位的重要性定义为该单位作为重要话语被主体提取的比例。虽然重要性与韵律参数之间没有较大的相关性，但对多个话语单元进行平滑处理后，功率参数的相关系数有所增加。5 .话语平滑给出了最大的相关性。与其他韵律参数相比，持续时间参数的相关性更大。

In order to investigate the efficiency of a stochastic FO model based on the clustering of accentual phrases, the accent type of accentual phrases in spoken sentences was estimated by the observed FO pattern and the model. Knowledge on accentual rules is introduced to reduce the potential accent type variations. The model was trained with 152 spoken lectures, and it was evaluated for other 15 spoken lectures in CSJ (Corpus of Spontaneous Japanese). It is shown that the introduction of accentual rules improves the performance of accent type estimation and the stochastic FO model is effectively constructed.Since speech data is not appropriate for quick scanning, the development of automatic summarization of lecture speech is expected. To realize the automatic summarization, the extraction of important sentences or words from speech data by hand was carried out. 17 subjects were asked to summarize 20 speech data from CSJ. It is confirmed that subjects are easy to agree on extracted key sentences for spoken lectures which have large prosodic inflections, based on the analysis of the agreement of extracted key sentences and the correlation between the extracted key sentences and FO variance.The relationship between the importance degree and prosodic parameters is analyzed for utterance units which are automatically segmented by more than 200ms pauses. The importance degree of utterance units is defined as the ratio that the unit is extracted as an important utterance by subjects. Although large correlation between the importance degree and prosodic parameters is not found, smoothing operation for several utterance units increased the correlation coefficients for power parameters. 5 utterance smoothing gives the largest correlation. Larger correlation as found for the duration parameter than other prosodic parameters.