A speech understanding framework that uses multiple language models and multiple understanding models

Abstract

The optimal combination of language model (LM) and language understanding model (LUM) varies depending on available training data and utterances to be handled. Usually, a lot of effort and time are needed to find the optimal combination. Instead, we have designed and developed a new framework that uses multiple LMs and LUMs to improve speech understanding accuracy under various situations. As one implementation of the framework, we have developed a method for selecting the most appropriate speech understanding result from several candidates. We use two LMs and three LUMs, and thus obtain six combinations of them. We empirically show that our method improves speech understanding accuracy. The performance of the oracle selection suggests further potential improvements in our system.