Autism spectrum disorder (ASD) is a medical condition originating as a growth-oriented illness that can affect children's social and cognitive skills, resulting in patterns of conduct, limited interests, difficulty communicating, and social engagement difficulties. Early identification of ASD can reduce the frequency and severity of the signs and symptoms. Deep learning (DL) is one of the most modern techniques that can be used for early ASD diagnosis or prevention of long-term consequences. As a result, automated methods for collecting objective measures of autism have been developed, and characteristics of the voice were in addition to being stated as distinct by healthcare providers, nevertheless have also shown promising efficacy in multiple research studies making use of deep learning techniques that are based on computerized exclusion comparing children with ASD to children with typical development (TD). To address these concerns, it suggests a feature recovery auto-encoder algorithm and an aggregate optimization procedure that may attain reliability for a large amount of unpolished data by utilizing a deep-learning-based technique that employs many different models to identify autism. Because of deep learning, the resulting results of these machine learning models were sent to a centralized database where a meta-classifier was developed to determine the approach that is more accurate in detecting autism. For feature extraction, four independent ASD patient datasets including over 600 records of affected children and adults were gathered from various repositories. The suggested approach accurately detects ASD in both children and adults at 99.45 and 99.84, respectively.