Deep learning techniques have played a pivotal role in advancing automatic speech recognition (ASR), historically integrating neural networks with hidden Markov models (HMMs). As speech is inherently dynamic, architectures rooted in recurrent neural networks (RNNs) have emerged as essential tools for capturing temporal dependencies and managing the complexities of natural speech. More recently, attention mechanisms have been incorporated into RNN-based encoder-decoder frameworks, providing both enhanced interpretability and more effective modeling of long-range contextual relationships. In this work, we present a hybrid ASR model that integrates multi-layer LSTMs, convolutional neural networks (CNNs), and attention, and we evaluate it on the TIMIT phoneme corpus, achieving a phoneme error rate (PER) of 23.6% on the test set. Beyond quantitative performance, we examine the impact of different attention mechanisms on training efficiency and final accuracy, finding that Bahdanau attention outperforms Luong attention in terms of convergence speed and recognition quality. Building on these insights, we propose optimization strategies aimed at accelerating learning while maintaining or enhancing final performance, contributing to more robust and efficient ASR solutions.