Ridwan Bello

Multi-Task Learning for African-Accented English Speech

Objective

This project explores multi-task learning models capable of simultaneously performing accent recognition, gender recognition, and age recognition from African-accented English speech data. The aim is to enhance both ASR performance and speaker demographic profiling in a unified framework.

Role & Contributions

My contributions include:

• Designing the multi-task learning architecture combining Wav2Vec2 embeddings with multiple classification heads.
• Preprocessing and labeling datasets with accent, gender, and age attributes.
• Implementing multi-task loss functions and training procedures.
• Evaluating performance across tasks using balanced accuracy and confusion matrices.

Technologies & Methodologies

• Python, PyTorch, Hugging Face Transformers
• Wav2Vec2, Multi-Task Neural Networks
• Scikit-learn, FastAPI

Outcomes & Impact

The multi-task model demonstrates strong accuracy across all tasks, with early results indicating enhanced generalization and a more robust speech recognition pipeline. Ongoing experiments aim to further optimize task balancing and explore fairness considerations.

Related Publications & Presentations

Currently in preparation for publication.