Residual Network Based Bidirectional Gated Recurrent Unit for Speech Recognition Using Speech Signals

Speech Recognition is challenging research ins speech processing field of research. Different method has available in previous research that attained high accuracy. However, accuracy of speech recognition is not increased because of irrelevant features in recognition stage. In this research, proposed a Residual Network based Bidirectional Gated Recurrent Unit (ResNet based Bi-GRU) method for speech recognition using Libri Speech dataset. The signals from data are pre-processed and feature extraction is performed through Mel Frequency Cepstral Coefficients (MFCC) that extract relevant features. Then, the extracted features are fed to classification stage which recognized the speech signals with high accuracy and less error rates. Performance of developed method is evaluated to metrics of Word Error Rate (%), Character Error Rate (CER), accuracy and f1-score. Developed technique obtained less WER of 4.55%, CER of 4.1%, accuracy of 96.71% and f1-score of 96.15% that is effective than existing methods like Recurrent Neural Network (RNN) –Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM).