Home
Repository Search
Listing
Academics - Research coordination office
R-RC -Acad
Admin-Research Repository
Engineering and Computer Science
Computer Science
Engineering
Mathematics
Languages
Arabic
Chinese
English
French
Persian
Urdu
German
Korean
Management Sciences
Economics
Governance and Public Policy
Management Sciences
Management Sciences Rawalpindi Campus
ORIC
Oric-Research
Social Sciences
Education
International Relations
Islamic thought & Culture
Media and Communication Studies
Pakistan Studies
Peace and Conflict Studies
Psychology
Content Details
Back to Department Listing
Title
A DEEP LEARNING METHOD FOR INNER SPEECH CLASSIFICATION USING EEG SIGNAL
Author(s)
Muhammad Ameer Hamza
Abstract
This thesis studies utilizing electroencephalography (EEG) signals the difficulties of subject-independent inner speech classification. Particularly for those with severe motor disabilities, inner speech the act of silently communicating to oneself offers a potential modality for Brain-Computer Interfaces (BCIs). Low signal-to-noise ratios and great inter-subject variability make deciphering inner speech from EEG difficult, nevertheless. Based on EEG data, this thesis evaluates several machine learning and deep learning models for inner speech classification. Particularly, a deep learning model, a Convolutional Neural Network (CNN) with triplet loss, is contrasted against conventional machine learning methods including Linear Support Vector Machine (SVM), More general SVM with various kernels, and LightGBM. Subject-independent framework with leave-one-subject-out cross-valuation on the Thinking Out Loud (TOL) dataset evaluates the models. Performance is evaluated with reference to accuracy, F1-score, precision, and recall. The CNN-based triplet network achieves the best average accuracy among other models, so the results show the promise of deep learning for subject-independent inner speech classification. Although the results imply that deep learning presents a viable path for future research, especially with bigger and more diverse datasets and advanced architectures, benefits over conventional approaches are minor. This work advances knowledge of the difficulties and possible solutions for creating strong, generally applicable inner voice BCIs.
Type
Thesis/Dissertation MS
Faculty
Engineering and Computer Science
Department
Computer Science
Language
English
Publication Date
Subject
Publisher
Contributor(s)
Format
Identifier
Source
Relation
Coverage
Rights
Category
Description
Attachment
Name
Timestamp
Action
32a243ab91.pdf
2025-09-02 09:30:35
Download