Efficient Access to Lecture Audio Archives through Spoken Language Processing

The paper firstly addresses the current state of speech recognition using the “Corpus of Spontaneous Japanese (CSJ)”. It is shown that the large-scale corpus had strong impact in training acoustic and language models considering morphological and pronunciation variations which are characteristic to spontaneous Japanese. Unsupervised adaptation of these models and the speaking rate is also effective, and we obtained word accuracy of 78.0%. Then, an intelligent archiving system of lectures based on automatic transcription and indexing is introduced. Transcriptions are automatically edited for improving readability, and key sentences are indexed based on statistically-derived discourse markers and topic words. Thus, we realize efficient browsing of lecture audio archives.