This paper addresses speaker adaptation of language model in large vocabulary spontaneous speech recognition. In spontaneous speech, the expression and pronunciation of words vary a lot depending on the speaker and topic. Therefore, we present unsupervised methods of language model adaptation to a specific speaker by (1) making direct use of the initial recognition result for generating an enhanced model, and (2) selecting similar texts for adaptation utterance by utterance. We also investigate the pronunciation variation modeling and its adaptation in the same framework. It is confirmed that all proposed adaptation methods and their combinations reduced the perplexity and word error rate in transcription of real lectures.
展开▼