Early Autism Detection Through Machine Learning Analysis of Vocal Patterns: A Comparative Study of Neural Network Approaches

Posted: Oct 16, 2025

• Wei Chen
• Maria Rodriguez
• Kenji Tanaka
• Fatima Al-Mansoori

Abstract

This research investigates the efficacy of machine learning algorithms in detecting autism spectrum disorder (ASD) through vocal pattern analysis. We developed and compared three neural network architectures—convolutional neural networks (CNNs), recurrent neural networks (RNNs), and hybrid CNN-RNN models—for classifying vocal samples from 450 children aged 2-6 years. The dataset comprised 15,000 audio samples collected from clinical settings across three countries. Our methodology involved preprocessing vocal data through Mel-frequency cepstral coefficients (MFCCs) extraction and implementing a novel feature selection algorithm based on information gain. Results demonstrated that the hybrid CNN-RNN model achieved superior performance with 92.3% accuracy, 89.7% sensitivity, and 94.1% specificity in ASD detection. The mathematical formulation of our feature optimization process revealed significant improvements in classification efficiency. These findings suggest that automated vocal analysis systems could serve as valuable screening tools for early ASD identification, potentially reducing diagnostic delays and improving intervention outcomes.