Online Temporal Language Model Adaptation for a Thai Broadcast News Transcription System

摘要: This paper investigates the effectiveness of online temporal language model adaptation when applied to a Thai broadcast news transcription task. Our scheme works as follow: first an initial is trained with available during development period. Then adapted over time more recent and articles deployment especially data from same period speech being recognized. We found that are closer in similar terms perplexity suitable for adaptation. The LMs better, both WER, than static LM only set data. Adaptation improved by 38.3% WER 7.1% relatively. Though, achieved less improvement, it still useful resource can be obtained automatically. Better pre-processing techniques selection based on text similarity could obtain further improvement this promising result.