Autism Speech Disorder (ASD) is a complicated neurodevelopmental condition that is linked to patterns in behavior and challenges in methyl social connection and communication along with repetitive repetitive and stereotyped action. Early and accurate identification using standardized tools of the condition is important in order to provide early, meaningful and appropriate intervention. The development of deep learning in the last couple of years has demonstrated immense capabilities of increasing the accuracy, as well as efficiency of ASD diagnosis through the use of multi source data, including neuroimaging, genetic make-up and behavioral profiles. The purpose of this paper is to systematically review different deep learning approaches used in the diagnosis of ASD as well as summarize their approaches, advantages, and disadvantages. We introduced convolutional neural networks (CNNs) for analyzing imaging features, recurrent neural networks (RNNs) for temporal behaviors, and multiple architectures integrated models for handling multimodal information. Furthermore, we present an overview of the feature extraction process, data pre processing techniques, and evaluation measures frequently used in the context of ASD research.