Automatic Speech Recognition (ASR) systems have significantly advanced with the integration of deep learning techniques, particularly neural networks and attention mechanisms; however, challenges remain in accurately modeling variable-length sequences and capturing complex temporal dependencies inherent in speech data. In this paper, we propose an enhanced end-to-end ASR model that incorporates a novel iterative attention mechanism, which loops through the attention layer multiple times with independent Dense and Dropout layers in each iteration. This design enables the model to focus on different aspects of the input sequence, effectively enhancing its ability to capture nuanced temporal patterns. Evaluated on the TIMIT dataset—a benchmark known for its comprehensive phonetic coverage—our model achieves a Phoneme Error Rate (PER) of 15.9%, outperforming existing neural network-based models. The results demonstrate that our iterative attention mechanism offers a more flexible and accurate solution for speech recognition tasks, addressing challenges associated with variable-length sequences and noisy environments, and contributing to the development of more robust ASR systems.