Employing Deep Learning to Create Speech Recognition Systems for Accented English

This research proposes a novel approach for developing an accented English speech recognition system using a deep learning model optimized by the African Buffalo Optimization (ABO) algorithm. Accented English speech recognition poses unique challenges, as variations in pronunciation and intonation across accents can significantly impact recognition accuracy. To address these challenges, the proposed model utilizes an ABO-LSTM framework, which combines the long short-term memory network’s capability to capture long-term dependencies with the ABO algorithm's strength in optimizing model parameters for enhanced accuracy. The methodology begins with data collection from a publicly available dataset, capturing English speech samples from speakers of various accents, including English, French, German, Italian, UK, and US. Audio preprocessing involves noise reduction through band-pass filtering to improve feature quality. Key speech characteristics are then extracted using Mel-Frequency Cepstral Coefficients (MFCC), which emulate human auditory perception and serve as input features for the model. The ABO algorithm optimizes the LSTM network by adjusting parameters such as learning rate, depth, and hidden units to minimize error rates. The results, demonstrated through confusion matrices and accuracy metrics, show significant improvement in accent classification accuracy of 95.4%, with the model successfully differentiating accents and handling linguistic variations in diverse contexts. This study contributes to speech recognition research by introducing a bio-inspired optimization method tailored for accented speech, with potential applications in cross-cultural communication, language learning, and automated transcription systems. The findings suggest that ABO-optimized LSTMs can effectively address accented speech recognition challenges, paving the way for more inclusive and adaptive speech recognition technologies.