Overview This project aims to develop a deep learning model for classifying basic directional commands in Ewe, a West African language. The goal is to improve navigation assistance for visually impaired individuals in linguistically diverse regions. Model and Approach
Implemented a MatchboxNet model from scratch Designed for audio classification tasks, specifically for recognizing spoken commands Key features:
Deep learning approach using audio recordings Language-specific training for Ewe Custom architecture optimized for the task
Dataset
Audio files of directional commands in Ewe Training Data: 200,000 samples Test Data: 120,000 samples
Project Pipeline
Data Cleaning and Visualization
Audio loading and resampling Silence removal Bandpass filtering Noise reduction Spectral subtraction Audio normalization
Model Training and Building
MatchboxNet architecture implementation Training process with data loading, loss function definition, and optimization
Future Work
Explore additional audio classification models Improve training data diversity Integrate advanced linguistic features Investigate transfer learning approaches Develop a user-friendly interface or mobile application
Conclusion This project demonstrates the potential of deep learning in addressing challenges faced by visually impaired individuals, particularly in linguistically diverse regions. Author ML_Wizard