Evaluating Unsupervised Language Model Adaptation Methods for Speaking Assessment
Shasha Xie, Lei Chen · 2013
In automated speech assessment, adaptation of language models (LMs) to test questions is important to achieve high recognition accuracy However, for large-scale language tests, the ordinary supervised training, which uses an expensive and time-consuming manual transcription process, is hard to utilize for LM adaptation. In this paper, several LM adaptation methods that require either no manual transcription process or just a small amount of transcriptions have been evaluated. Our experiments suggest that these LM adaptation methods can allow us to obtain considerable recognition accuracy gain with no or low human transcription cost. Index Terms: language model adaptation, unsupervised training, Web as a corpus