Human-like Spoken Dialog System

在本研究中，我们研究了对话节奏对用户舒适度和可靠性的影响，并在此基础上提出了一个构建语音界面的新框架。虽然我们确认了用户对话的满意度和流畅性的提高，但我们还没有达到人与人之间对话的自然程度。为了实现这一点，我们改进了生成有节奏对话框的模型，将其重新实现到系统中，并提高了处理速度。结果表明，与之前的系统相比，我们获得了更好的人类相似性和可靠性，但由于系统的时间识别性能水平、语音识别和语言理解性能水平不完善，我们无法达到与人类相似的评价。

In this research we study how dialog rhythm influences user's comfort and reliability and propose a new framework for building spoken interfaces based on this framework. Although we confirmed user's increased satisfaction and smoothness of conversation, we have not reached the level of naturalness of human to human dialog.To achieve this we have improved our model for generating rhythmical dialogs, re-implemented it into the system and increased processing speed.In result, we achieved better human-likeness and reliability comparing to the previous system, but we could not reach evaluation similar to human's as our system's timing recognition performance level, voice recognition and language understanding performance levels were imperfect.