Automatic Transcription of Meetings Using Topic-oriented Language Model Adaptation

This paper presents an automatic speech recognition (ASR) system dedicated for meetings of the National Congress of Japan. The distinctive features of the congressional meeting speech are wide distribution and frequent change of topics. For more accurate transcription, such topics should be emphasized in a language model one after another. Therefore, we introduce two approaches for topic adaptation: PLSA-based approach and trigger-based approach. The PLSA-based adaptation is performed turn by turn to emphasize topics in individual pair of a question and answer. Since topics are treated in a probabilistic manner, robust adaptation is realized. On the other hand, the trigger-based adaptation stresses relevant words to the word history, thus long-distance context can be reflected into a language model. These two approaches were evaluated on real meetings of the Congress, and significant improvement of perplexity was obtained by both approaches. We also compared their effects on reduction of word error rates.