Accent Classification with Wav2Vec2
Objective
This project aimed to develop a speech accent classification system that accurately identifies various African English accents. The focus was on leveraging pre-trained deep learning models to handle accented speech more effectively in low-resource settings.
Role & Contributions
I led the design and implementation of the classification pipeline, which involved:
- Curating and preprocessing a diverse dataset of African English accents.
- Extracting MFCC features and Wav2Vec2 embeddings.
- Building and training classification models (MLP, SVM).
- Evaluating model performance using cross-validation and tuning hyperparameters.
Technologies & Methodologies
- Python, PyTorch, Hugging Face Transformers
- Wav2Vec2, MFCC feature extraction
- Scikit-learn (SVM, MLP classifiers)
- 5-fold cross-validation, learning rate scheduling
Outcomes & Impact
The system achieved an accuracy of 82% across multiple accents, demonstrating the potential of fine-tuned pre-trained models for accent classification. Results are to be presented at the 2026 African AI Conference, and a manuscript is under preparation for submission to IEEE TASLP.
Visuals
MFCC + MLP Model Training/Validation Curve
Confusion Matrix of Accent Classification
Related Publications & Presentations
- Bello, R. "Accent-Aware Speech Recognition Using Wav2Vec2." IEEE ICASSP, 2024. View Paper
- Presentation at African AI Conference 2024 – View Slides