Abstract—Automatic speech recognition (ASR) plays a vital role in various domains, improving search engines, aiding healthcare with medical reporting and diagnosis, enhancing service delivery, and facilitating effective communication in service providers. This paper introduces FNNRA (Flexible Neural Network with Recursive Architecture), a novel method aimed at addressing overfitting issues in environments with limited training datasets in the field of automatic speech recognition (ASR). FNNRA utilizes a sophisticated architecture to extract and analyze important data features while maintaining data integrity through deep network layers. Theoretical and practical evaluations demonstrate FNNRA's ability to handle speaker variations, effectively train with limited datasets, and extend its applicability beyond speech recognition. The method is evaluated on established datasets like CallHome, TIMIT, and FarsDAT, showcasing its adaptability and efficacy across different data contexts. Comparative analysis with leading speech recognition methods reveals FNNRA's superior performance, achieving significant reductions in phoneme recognition errors by approximately 7.88%. This research sets a strong foundation for future advancements in the field and underscores FNNRA's potential in enhancing recognition systems, warranting further investigation.