Emotion detection in children's speech using linguistic, paralinguistic and breathing features with AI methods
Summary
The field of Human-Computer Interaction is increasingly emphasizing the need for computers to react in a more human-like manner. To achieve this, computers must be able to comprehend the emotional context of conversations. However, as young children are also interacting with technology nowadays, the absence of emotional and conversational experience with children in current models could be problematic. To address this, the Dutch children's emotional speech and breathing dataset has been added to the emotional databases in the affective computing field. With this dataset, various types of analyses can been conducted, including speaker diarization, audio transcription, linguistic and paralinguistic feature analysis, and breathing analysis. The findings indicate that the automatic segments and transcription are not suitable for further analysis in their current state. However, manual transcriptions, with or without augmentation, result in relatively good performance by the state-of-the-art linguistic models used for emotion recognition. In contrast, the paralinguistic models show relatively poor performance, which could be due to the noise and sparsity of the data or the lack of children's data used to develop the feature sets. Furthermore, the breathing data is found to be correlated with the Self-Assessment Manikin (SAM) scores, suggesting that children's breathing data could enhance the current emotion recognition models.