Sound Source Localization for Autistic Children's Session Recordings

This work addresses the problem of localizing simulated sound sources for the Autism Diagnostic Observation Schedule 2^nd edition (ADOS-2) observation room. The main challenge stems from the unconventional, nonuniform infrastructure-installed microphone array that prevents the adoption of the conventional beamforming approaches. The proposed deep neural network (DNN)-based source localization approach is formulated as a classification problem, where the sound source location is classified into one of four zones within the ADOS-2 room. Two architectures were introduced based on a bidirectional long short-term memory (BiLSTM) network and a hybrid BiLSTM with a transformer encoder. The classification accuracy of 85-89% was demonstrated in diverse acoustic environments using various microphone array configurations. It was shown that the proposed approach could achieve an efficient localization performance in clinical settings, indicating its potential applications in autism diagnosis and treatment.Clinical relevance-Speech-based sound source localization (SSL) in autism evaluation sessions can provide valuable insights into children's spatial behavior and interaction patterns. Additionally, it has the potential to improve the quality of speech recordings, thereby supporting more accurate autism diagnosis and intervention strategies.