EEG-BASED IMAGINED SPEECH RECOGNITION USING CONVOLUTION NEURAL NETWORK MODEL

Abstract

EEG based BCI is intended to provide a new medium of communication to people who are unable to speak due to motor disability or some neurological disorder. Electroencephalography (EEG) signals which are non-invasive are one of the most used brain signals in BCI applications. They are usually contaminated with noise which poses some challenges in accurate recognition. Machine learning (ML)-based algorithms have had little success decoding imagined speech, mostly due to the weaker and more unpredictable brain impulses compared to overt speech. Convolutional neural networks (CNNs) and deep learning (DL) have revolutionised computer vision in recent years, outperforming conventional machine learning (ML)-based algorithms in pattern identification. This project mainly deals with analysis of EEG signals and investigating the feasibility of recognizing imagined speech from EEG signals using various time -frequency methods including Morlet Wavelet transform, Short-Time Fourier Transform (STFT), Stockwell Transforms, to extract discriminative features from EEG signals during imagined speech tasks. To see the geographic distribution of brain activity, topographic maps of EEG power across several frequency bands are also created. To classify imagined speech, a Convolutional Neural Network (CNN) model is trained using the collected features and Topographic maps pictures as inputs. The results mainly indicate that on giving the images to CNN model, topographic maps provide better result with accuracy as compared to other time-frequency representation methods like Morlet, Stockwell and Short-time Fourier transform. The CNN model used in our proposed model outperformed other classification models like K-Nearest Neighbours (KNN) and Support Vector Machines (SVM).