Computational Paralinguistics involves analyzing the expressions of users to obtain information about individual characteristics and different mental states. To date, most research on paralinguistic information has targeted primarily adult speech and, to a lesser extent, child speech, even though children are potential beneficiaries of computers with speech-based interfaces (e.g., for educational applications and entertainment purposes). Current children’s speech corpora have been designed primarily for tasks not related to paralinguistic analysis, such as speech recognition or the study of acoustic properties. Few of these corpora contain genuine interactions that comprise annotations of paralinguistic phenomena. Therefore, to create new emotional models that help develop more advanced speech-based interactive systems or models designed for children, new corpora of child-computer interactions that complement the existing corpora are needed.
To contribute to the analysis of paralinguistic information in children’s speech we created the IESC-Child. This corpus consists of 2,093 minutes of audio recordings (34.88 hours) divided into 19,283 speech segments of the interactions of 174 children (both sexes) between the ages of 6 and 11 years who are speaking in Mexican Spanish. The recordings were manually segmented and transcribed by seven human annotators. The native language of each annotator was Spanish. IESC-Child contains two types of paralinguistic information labels: emotions (anger, disgust, fear, happiness, sadness, and surprise) and attitudes (confident, uncertain, apathetic and enthusiastic). For the sake of simplicity, we obtained the final classes for the two labeling categories by plurality vote across all the annotators.
To produce different affective reactions in the participants, we used a Wizard of Oz (WoZ) scenario where children participated in interactive sessions with two Lego Mindstorms behaving either collaboratively or noncollaboratively. This corpus can facilitate the study of affective reactions in speech communication during child-computer interaction in Spanish and the creation of models for the recognition of acoustic paralinguistic information. The Interactive Emotional Children’s Speech Corpus (IESC-Child) can be a valuable resource for researchers given that offers some advantages over currently available databases such as its size, naturalness, and the type of information annotated. IESC-Child is publicly available for further analysis of the collected data and further research.