ISCA Archive SpeechProsody 2022
ISCA Archive SpeechProsody 2022

Acoustic discriminability of unconscious laughter and scream during game-play

Takuto Matsuda, Yoshiko Arimoto

For the growing demand to make social signals available to various systems, such as laughter for clinical treatment or emergent screams for security reasons, it is essential to detect those signals as appropriate social functions. This study demonstrates the acoustic discriminability of laughter and scream by conducting three machine-learning-based classification experiments and features selection experiments based on logistic regression analysis, using classical acoustic features. The result of the speaker-and-corpus-closed experiment revealed that the models acoustically discriminate laughter from screams by yielding high accuracies at 93.52% (DNN) and 95.54% (SVM). Moreover, the result of the leave-four-speaker-out cross-validation (LFOCV) revealed that our model can correctly classify laughter and screams regardless of the speakers by exhibiting only approximately 1% lower accuracies than the result of the speaker-closed model. The results of the corpus-open experiment exhibited approximately 5% and 7.8% lower accuracies for the DNN and SVM models, respectively, than those of the corpus-closed experiment. However, our model can still classify laughter and screams in the different recording conditions at approximately 88% accuracy. Finally, the result of logistic regression analysis showed that the harmonics-to-noise ratio was the most contributed acoustic feature to discriminate laughter from screams.

doi: 10.21437/SpeechProsody.2022-117

Cite as: Matsuda, T., Arimoto, Y. (2022) Acoustic discriminability of unconscious laughter and scream during game-play. Proc. Speech Prosody 2022, 575-579, doi: 10.21437/SpeechProsody.2022-117

  author={Takuto Matsuda and Yoshiko Arimoto},
  title={{Acoustic discriminability of unconscious laughter and scream during game-play}},
  booktitle={Proc. Speech Prosody 2022},