A Hybrid Architecture Combining CNN, LSTM, and Attention Mechanisms for Automatic Speech Recognition

This research proposes a robust sequence-tosequence model architecture for speech recognition tasks, incorporating convolutional and recurrent neural networks with attention mechanisms. The encoder utilizes a hybrid approach of residual convolutional blocks (RCNN) and bidirectional Long Short-Term Memory (BLSTM) networks to extract hierarchical temporal and spectral features from input sequences efficiently. The decoder employs an attention-augmented LSTM with Luong Attention, dynamically focusing on relevant parts of the encoded representation during sequence generation. To enhance generalization and robustness, we integrate data augmentation techniques such as phase randomization, time-frequency masking in the preprocessing pipeline. Features are further refined using Mel-Frequency Cepstral Coefficients (MFCC) and Filter Banks (FBANK) combined with delta and delta-delta coefficients. Beam Search decoding with a width of 10 is used during inference to improve the accuracy of predictions. We evaluate the model on the TIMIT dataset and achieve a Phoneme Error Rate (PER) of $\mathbf{1 6. 7 \%}$, demonstrating the effectiveness of the proposed architecture and preprocessing strategies in capturing complex phoneme patterns and improving recognition performance.