CHERISH: A Corpus for CHildren Emotion Recognition In SpeecH

Resumen

Emotion recognition in children remains significantly less explored than in adults, largely due to the limited availability of annotated data. To address this gap, we introduce CHERISH, a multimodal corpus for emotion recognition in Spanish-speaking children. Our data collection methodology includes various sources such as speech, speech transcriptions, behavioral descriptions provided by a human observer, and personality traits obtained through the Children's Personality Questionnaire (CPQ). Each of these modalities contributes key information to enhance children emotion recognition: speech provides insights into vocal expression, transcriptions reflect the semantic content of speech, behavioral descriptions offer context on body language, and personality since it influences how children express and regulate their emotions. We aim to contribute research on emotion recognition in children, which could enhance the development of more robust models that can be applied in education, healthcare, and child-assistive technologies. Additionally, we present baseline results for the emotion recognition task under two experimental conditions, establishing a foundation for future comparative studies.