Assessing the effectiveness of diarization algorithms in costa rican children-adult speech according to age group and gender

Speaker diarization is the task of automatically identifying speaker identities and detecting their speaking times in an audio recording. Several algorithms have shown improvements in the performance of this task during the past years. However, it still has performance challenges in interaction scenarios, such as between a child and adult, where interruptions, fillers, laughs and other elements may affect the detection and clustering of the segments. In this work, we perform an exploratory study with two diarization algorithms in children-adult interactions within a recording studio and assess the effectiveness of the algorithms in different age groups and genders. All participants are native Costa Rican Spanish speakers. The children have ages between 3 to 14 years, and the interaction combines guided repetition of words or short phrases, as well as natural speech. The results demonstrate how the age affects the diarization performance, both in cluster purity and speaker purity, in a direct but non-linear fashion.