Evaluation of Spoken Dialogue System that uses Utterance Timing to Interpret User Utterances

Abstract

In real environments where automatic speech recognition (ASR) performance is not always high, an alternative strategy to relying on ASR results is required to achieve a robust speech interaction.We construct a spoken dialogue system that can enter into the enumeration subdialogue, in which the utterance timing as well as the ASR results are used to interpret user utterances. Since the utterance timing can be obtained more reliably than the ASR results, the subdialogue leads to achieve a more robust interaction even in situations with a low ASR performance. We conducted an experiment with 31 participants. The results showed that our system achieved a higher task completion rate than a baseline system that uses only the ASR results when the ASR accuracy was not high.