A Novel Deep Learning-based Brain Computer Interface for Automatic Speech Recognition using Multistyle Training-based Auditory Generative Adversarial Network

Abstract

Brain-Computer Interfaces (BCIs) provide significant potential in neuroscience and rehabilitation. Such systems allow individuals with neurological impairments to communicate thoughts externally without the need for physical movement. Despite this promise, no standardized frameworks currently exist for validating the amount of speech produced by BCI-based synthesizers. Decoding speech remains difficult because neural signals are inherently weak and show greater variability than spoken output. Conventional approaches often perform poorly in noisy environments, limiting effectiveness. Variations in accents and regional speech patterns also reduce the accuracy of automatic speech recognition when analyzing speech signals. Recently, deep learning models have transformed computer vision and demonstrated superior pattern recognition capabilities compared to traditional algorithms. Therefore, this research aims to introduce a novel deep learning-based BCI for automatic speech recognition using speech signals. Initially, the required speech signals are gathered from public data sources. Later, the speech signals are forwarded to the designed Multistyle Training-based Auditory Generative Adversarial Network (MT-AGAN) model to execute the speech recognition process. Finally, the effectiveness of the designed method is analyzed by performing diverse comparative analyses over existing methods.