ogeelekwa/Nigerian-Offline-Command-Recognizer

This project demonstrates an offline speech recognition pipeline designed for low-resource African languages on Android devices. It was developed as part of **awadoc**, an AI-powered health assistant for Nigerian users with limited internet connectivity # Offline Nigerian Voice Command Recognizer **An offline-first voice command recognition system for Nigerian languages (Igbo, Hausa, Yoruba)** This project demonstrates an offline speech recognition pipeline designed for low-resource African languages on Android devices. It was developed as part of **Awadoc**, an AI-powered health assistant for Nigerian users with limited internet connectivity. ## Key Features - ✅ **Fully offline** - No cloud APIs required, runs entirely on-device - ✅ **3 Nigerian languages** - Igbo, Hausa, Yoruba support - ✅ **Closed-grammar command recognition** - ~30 fixed commands per language - ✅ **Android deployment** - TFLite model optimized for low-end devices - ✅ **Low-resource optimization** - INT8 quantized models (~5MB each) - ✅ **Real-time inference** - <100ms latency on Android ## Technical Stack | Component | Technology | |-----------|-----------| | Model Architecture | wav2vec2-xls-r / Custom KWS CNN | | Training | PyTorch + Hugging Face Transformers | | Deployment | TensorFlow Lite (TFLite) | | Quantization | INT8 (quantization-aware training) | | Android | Android Studio + TFLite Interpreter | | Languages | Igbo, Hausa, Yoruba (low-resource African languages) |