Japanese Speech Intelligent Recognition Using Convolutional Neural Network with Bidirectional Gated Recurrent Unit

In recent years, Speech recognition is the technology that converts spoken language into text, whereas Japanese speech recognition face different challenges due to the language's complex morphology. To overcome these challenges in this research improved Convolutional Neural Network (CNN)- Bidirectional Gated Recurrent Unit (BiGRU) namely CNN- BiGRU proposed for recognizing the imagined speech. Initially, the input data is collected from the Kara One database, which is an ElectroEncephaloGram (EEG)-based dataset designed for imagined speech recognition. Then, the preprocessing is performed using notch filtering for noise removal and Independent Component Analysis (ICA) for artifact removal. Further, Fast Fourier Transform (FFT) is utilized for extracting the frequency-domain features and Linear Discriminant Analysis (LDA) used for discriminative features. Finally, the CNN extracts spatial features from EEG signals and identifies relevant speech-related brain activity. The Bi-GRU captures temporal dependencies in both forward and backward directions, effectively processing the sequential patterns of imagined speech to recognize Japanese speech. The proposed Improved CNN-BiGRU achieved better results in terms of accuracy with 78% compared to the existing CNN-Long-Short Term Memory (LSTM).