Gaming Corpus for Studying Social Screams.

Screams in everyday conversation, rather than in emergencies, are considered as nonverbal behavior that makes our speech communication rich and expressive. This paper focuses on screams in conversation. Identification of screams in existing spoken dialog corpora revealed that these corpora contained only a small number of screams, so not adequate for the investigation of the screams. In order to obtain more screams that naturally occur in conversation, we recorded dialogs while playing highly action-oriented games. Following to our criteria to identify screams, 1437 screams were detected from the whole recordings. The screams in our corpus were 12 times more frequent than the existing gaming corpus. As for the number of screams per minute, a strong positive correlation was observed between two speakers of the same pair, suggesting that the interlocutors produced screams not purely spontaneously, but tried to get the screaming behavior closer to the other person. Results of the acoustic analysis showed that the typical scream is produced in 60-140 mel higher and 8 dB louder voice than typical normal speech.