Prosody-based detection of the context of backchannel responses

Current spoken dialogue systems lack positive feedback such as backchannels, which are common in human-human conversations. To develop more natural human-computer interfaces, the investigation of backchannel-responses are indispensable. In this paper, we propose a method for detecting the precise timing for backchannel responses in Japanese and aim at incorporating such method in future spoken dialogue systems. The proposed method is based on machine learning technique with a variety of prosodic features. It is shown to be effective in automatically deriving rules for detecting the contexts of backchannels. The performance of our method is considerably better than previous methods.