A comparative study on extractive speech summarization of broadcast news and parliamentary meeting speech

We carry out a comprehensive study of acous-tic/prosodic, linguistic and structural features for speech summarization, contrasting two genres of speech, namely Mandarin Broadcast News and Cantonese Parliamentary Speech. We find that structural features are superior to acoustic and lexical features when summarizing broadcast news because of the fact that in the same Mandarin broadcast program, the distribution and flow of summary utterances are relatively consistent. We use different machine learning algorithms to construct the binary-class summarizers to select the best features for extractive summarization, and obtain state-of-the-art performances: ROUGE-L F-measure of 0.682 for Mandarin Broadcast News, and 0.737 for Cantonese Parliamentary Meeting Speech. In the case of Parliamentary Meeting Speech summarization, we show that our summarizer performed surprisingly well ROUGE-L F-measure of 0.729 by using ASR transcription despite the character error rate of 27%. We also discover that the different choices of algorithms almost do not affect the consistency of our findings.