Ridwan Bello

Bias & Fairness in Speech Recognition

Objective

This project investigates bias and fairness challenges in automatic speech recognition (ASR) systems, particularly when processing African-accented English speech. The goal is to identify disparities in ASR performance across demographic groups and develop fairness-aware training strategies to mitigate bias.

Role & Contributions

I am leading the experimental design and fairness evaluation strategy, including:

• Dataset auditing to analyze representation across accents, genders, and age groups.
• Defining fairness metrics (e.g., error rate disparity, equalized odds).
• Training and comparing baseline vs. fairness-aware ASR models.
• Documenting findings and proposing improvements to mitigate bias.

Technologies & Methodologies

• Python, TensorFlow, Hugging Face Transformers
• Fairness libraries: Fairlearn, AIF360
• Statistical analysis & visualization: Pandas, Seaborn, Matplotlib
• ASR models: Wav2Vec2, Whisper

Outcomes & Impact

This is an ongoing project. Initial results indicate notable disparities in word error rates (WER) between different accent groups, reinforcing the need for fairness-aware training. The anticipated impact includes improved fairness in ASR systems and contributions to the broader field of ethical AI.

Related Publications & Presentations

Publications and presentations will be added upon project completion.