Desarrollo de un módulo para el pre-diagnóstico de problemas de lenguaje en niños mediante un robot social

Abstract

This paper presents an AI module integrated into a social robot to support the early pre-diagnosis of childhood language difficulties through playful, non-invasive interactions. The system adopts a modular architecture comprising audio input and preprocessing, automatic speech recognition using fine-tuned Whisper for Spanish child speech, and linguistic analytics with automated reporting. A proof-of-concept was conducted in naturalistic settings (classroom and home) with two participants (ages 4 and 6). The fine-tuned model reduced WER from 16% to 11% compared with the base model. Analytics estimated a modal utterance length of 4–5 words and an average type–token ratio (TTR) of 0.42 (SD=0.07), with lower lexical diversity in the participant with suspected impairment. The classifier flagged 15% of records as “possible indications.” Response latency averaged 8.5 s for children with difficulties versus 4.2 s otherwise. Event-level confusion metrics were: Sensitivity = 0.65; Specificity ≈ 0.82; PPV ≈ 0.70; NPV ≈ 0.78. Children accepted the robot well, which fostered engagement. Findings indicate the approach is feasible as supportive screening in resource-constrained settings; it does not replace professional assessment. Future work includes enlarging the sample, adding multimodal signals, and optimizing deployment.