口調から言葉の尺度を推定する会話システム

抄録

Conventional conversational system technologies use speech to estimate the speaker’s intentions. However, the problem is that it is not possible to estimate the other party’s intention from various nonverbal information as in human conversation. In this paper, to use a neural network to learn about the relationship between tone of voice and probability word. Using this neural network, we estimated the speaker’s intention by analyzing the verbal range from tone. We proposed a Cue bar task to check the performance of this conversation system. We defined an evaluation function based on the total amount of movement. The results showed that the median of the evaluation function was less than 0.3. In other words, although the accuracy was low, it was confirmed that the speaker’s intention could be estimated. As a problem, it was suggested that the task needs to take into account the information transfer between humans and robots.