Dofadar, Dibyo FabianNeeha, Nuhash KabirRahman, NuzhatIslam, Imtela2026-01-202026-01-2020252025-06ID 21301025ID 21301538ID 21341018http://hdl.handle.net/10361/27465Cataloged from PDF version of thesis.Includes bibliographical references (pages 41-43).This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science and Engineering, 2025.Deep models have improved speech emotion recognition, yet overlapping emotions and speaker variability still limit practical deployment. The guiding hypothesis of this paper is that a useful solution must clearly tackle overlapping and blended emotions, along with enhancing classification strength. These issues are addressed with a multi-branch ensemble approach and enhanced with a fuzzy post-processing phase aimed at revealing secondary emotion memberships and emphasizing unclear, overlapping statements; speaker-adversarial training was added as a supportive strategy to minimize confounding factors. evaluation on augmented benchmark corpora demonstrates competitive accuracy and improved handling of label overlap. This dual emphasis on accuracy and interpretability provides a principled way to surface uncertainty in emotional input and aims to provide a foundation for SER systems to be more robust for real-world use.51 pagesenBRAC University theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission.Emotion recognitionSpeech analysisNatural language processingDeep learningVocal expressions detectionMachine responsivenessNatural interactionsNatural language processing (Computer science).Automatic speech recognition.Pattern recognition systems.Emotions--Identification--Data processing.Speech processing systems.Context-aware computing.Emotions--Computer simulation.Thesis