Okemwag/project-recordings

Accent-aware Kiswahili isolated-word speech recognition prototype using MFCC features, SVM, and ANN models. # Accent-Aware Kiswahili Speech Recognition System An isolated-word Kiswahili speech recognition system that accounts for regional accent variation across three Kenyan accent groups — **Coastal**, **Nairobi**, and **Upcountry**. The system uses MFCC-based feature extraction (with delta and delta-delta coefficients) fed into classical ML models (SVM and ANN), making it accessible for research on modest hardware without GPU infrastructure. > **Scope**: This is an *isolated-word* recognizer. It classifies single spoken Kiswahili words. It is **not** a conversational or continuous speech ASR system. --- ## Table of Contents 1. Background 2. System Architecture 3. Project Structure 4. Supported Accents 5. Prerequisites 6. Installation 7. Data Preparation 8. Configuration 9. Training 10. Evaluation 11. Inference 12. Testing 13. Known Limitations 14. Contributing 15. License --- ## Background Kiswahili is spoken by over 200 million people across East and Central Africa, yet it remains underrepresented in speech technology research. A key challenge is **accent variability**: speakers from the Kenyan coast, Nairobi, and upcountry regions pronounce the same words with measurably different phonetic patterns. This project addresses that gap by: - Building a balanced, multi-accent dataset of isolated Kiswahili words - Extracting robust MFCC + Δ + ΔΔ features that capture spectral and temporal dynamics - Training and comparing SVM (RBF kernel) and ANN (feedforward) classifiers - Reporting per-accent accuracy breakdowns to surface model bias The system is designed for telecom-adjacent use cases such as Interactive Voice Response (IVR) systems deployed in East Africa. --- ## System Architecture ``` Audio Input (WAV file or microphone) │ ▼ ┌─────────────────────────────┐ │ Preprocessing Module │ │ • Resample → 16 kHz │ │ • Spectral noise reduction │ │ • Silence trimming │ │ • Amplitude normalization │ └────────────┬────────────────┘ │ ▼ …