Automatic Speech Recognition Adaptation to the IoT Domain Dialogue System

摘要

The paper addresses the issue of error correction of the domain-specific output from the cloud Automatic Speech Recognition (ASR) system. The research and the solution were built for the Internet of Things (IoT) data collected in the process of applying voice control over home appliances. We describe an ASR post-processing module that reduces the word error rate (WER) of ASR hypotheses and consequently improves prediction of users' intentions by the Natural Language Understanding (NLU) module. The study also compares three various English proficiency level groups of speakers and makes observations about differences of recognition accuracy ratio of users' speech by a dialogue system.