Cache-based Language Model Adaptation using Visual Attention for ASR in Meeting Scenarios

摘要

In a typical group meeting involving discussion and collaboration, people look at one another, at shared information resources such as presentation material, and also at nothing in particular. In this work we investigate whether the knowledge of what a person is looking at may improve the performance of Automatic Speech Recognition (ASR). A framework for cache Language Model (LM) adaptation is proposed with the cache based on a person's Visual Attention (VA) sequence. The framework attempts to measure the appropriateness of adaptation from VA sequence characteristics. Evaluation on the AMI Meeting corpus data shows reduced LM perplexity. This work demonstrates the potential for cache-based LM adaptation using VA information in large vocabulary ASR deployed in meeting scenarios.