[3Xin2] Poster session 1

[3Xin2-09] Building Automatic Laugh Detector and Mental Disease Classifier Using Mental Disease Dialog Corpus

〇sho katsuki¹, Hironobu Nakamura², Taishiro Kishimoto³, Yoshinobu Kano¹ (1.ShizuokaUnivercity, 2.Tokyo Medical and Dental University, 3.Keio University)

Keywords:Mental Health, Disease Type Prediction, Laugh Detection

The UNDERPIN large-scale mental disorder conversation corpus we have constructed consists of various diagnostic results of subjects and their conversation recordings, all of which have been transcribed manually. Furthermore, a portion of these have been annotated with detailed phonetic and linguistic annotations by hand. However, considering the practical use for future automated diagnostic support, automatic annotation is desirable. In our previous research, statistics on laughter as a vocal feature have been suggested to be an important clue for diagnosis regardless of the disorder, yet this too relies on manual annotation. In this study, we have annotated laughter in the UNDERPIN corpus and fine-tuned the speech recognition system Whisper to construct an automatic laughter detection system. The detection performance of laughter exceeded 90%, achieving a high level of practical performance. Using the constructed automatic laughter detection system, we annotated dialogue text data in the UNDERPIN corpus and conducted disease classification experiments. The results showed performance comparable to that using manual annotations, demonstrating a practical level of performance for applications.