Integrating Attention Mechanisms with Bidirectional Long Short-term Memory Recurrent Neural Networks for Improved Speech Recognition

Speech-to-text technology is indispensable for converting spoken words into text, facilitating easy storage and retrieval. This process involves several critical stages, progressing from a basic model through signal pre-processing, feature extraction, feature selection, and modeling. There has been extensive literature on improving speech recognition, however, challenges persist, particularly in addressing word error rates, accuracy, latency, and computational efficiency in continuous input streams in a noisy environment. This research systematically evaluates the effectiveness of recurrent neural networks (RNNs), long short-term memory neural networks (LSTMs), gated recurrent units (GRUs), bi-directional long short-term memory (Bi-LSTM), and bi-directional long short-term memory (Bi-LSTM) with attention mechanism networks. It further proposes integrating attention mechanisms with a Bidirectional Long Short Memory Recurrent neural network to enable the model to focus on the crucial parts of the input sequence thus increasing computational efficiency. Experimental results, based on the Libri speech and TIMIT training dataset, demonstrate promising outcomes. The proposed model exhibits superior performance, achieving a word error rate of 1.6%, and an accuracy of 98.4% on the training dataset. As a future avenue of exploration, the study proposes evaluating the use of transformer models to reduce character error rates further and enhance accuracy in continuous input streams.