Identification of Target Speech Utterances from Real Public Conversation

摘要

We are developing a conversation support system that can estimate the smooth progress of human-to-human conversation. When the system senses there has been little progress in the conversation, it attempts to provide a topic to lead a smoother discussion and good atmosphere. The conversation atmosphere is estimated using the fundamental frequency (FO) and sound power (SP). In its practical use, the following problems occur: 1. Ambient noises, especially nonstationary speech signals of a person behind the target speaker, decrease the conversation-atmosphere estimation rate. It is difficult to cancel this speech noise, even when using current noise cancelling methods. 2. Laughter utterances in which acoustic characteristics are quite different from usual speech utterances are often seen in daily conversation, which causes a decrease in the conversation-atmosphere estimation performance. In this paper, we propose an identification method for target speech utterances from ambient speech noises or laughter utterances using the standard deviation value of SP and Mel-Frequency Cepstral Coefficients (MFCC).