Enhancing Performance of Noise-Robust Gujarati Language ASR Utilizing the Hybrid Acoustic Model and Combined MFCC + GTCC Feature

The study introduces an enhanced method for improving the accuracy and performance of End-to-End Automatic Speech Recognition (ASR) systems. This involves combining Gammatone Frequency Cepstral Coefficient (GTCC) and Mel Frequency Cepstral Coefficient (MFCC) features with a hybrid CNN-BiGRU model. MFCC and GTCC features capture temporal and spectral aspects of speech, while the hybrid architecture enables effective local and global context modelling. The proposed approach is evaluated using a low-resource Gujarati multi-person speech dataset, incorporating clean and noisy conditions via added white noise. Results demonstrate a 4.6% reduction in Word Error Rate (WER) for clean speech and a significant 7.83% reduction in WER for noisy speech, compared to baseline MFCC with greedy decoding. This method exhibits potential for enhancing ASR systems, making them more reliable and accurate for real-world applications necessitating precise speech-to-text conversion.