Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that significantly affects speech, communication, and social interaction. Early intervention is essential, yet most existing speech training systems are limited to English, use very restricted vocabularies, and are not adapted for Arabic-speaking children. This study proposes an AI-based pronunciation training system designed specifically for Arabic-speaking children with ASD. The system integrates Text-to-Speech (TTS) for generating clear reference pronunciations and Whisper-based Automatic Speech Recognition (ASR) for transcribing and evaluating the child’s speech. Due to the lack of publicly available Arabic ASD speech datasets, synthetic data augmentation was used to improve robustness. The system evaluates pronunciation using two main metrics: Accuracy (exact match) and Similarity (normalized edit distance), enabling more flexible and encouraging feedback. A test set of 50 Modern Standard Arabic words was used for evaluation. Results showed an overall word accuracy of 76.5%, similarity of 85.2%, Word Error Rate of 23.5%, Character Error Rate of 14.8%, and Mean Opinion Score of 4.2/5. The findings indicate that the proposed system can reliably detect near-correct pronunciations and provide positive reinforcement even when strict accuracy is low. This suggests its potential as a supportive tool for incremental speech development in children with ASD, especially in Arabic-speaking environments.