Language model adaptation based on correction information for interactive speech transcription
Jia Duan, Xiangdong Wang, Yuzhuo Ma, Yang Yang, Hong Liu, Yueliang Qian · 2016
Aiming at language model (LM) adaptation for interactive speech transcription, this paper proposes a topic-based adaptation method using users' correction information. To infer the topic for each utterance in continuous speech, this method uses the correction information of history utterances adjacent to the current one. Perplexity is calculated for topic inference. Topic-related LMs are interpolated with background LM to obtain adapted LMs. Each utterance is transcribed using the adapted model. This method is a supervised adaptation method which is believed to outperform the unsupervised approaches widely used in current speech recognition applications, since it uses the history of user correction. And this method is an online adaptation method for it adapts models before transcribing each utterance. Besides, utterance-level adaptation makes the adapted model much more precise for each utterance. Experimental results have shown that this method raises the average recognition accuracy rates by 2-6 percentage points.