abdulmunimjemal/SagalNet

SagalNet is a robust end-to-end Machine Learning pipeline for converting spoken Afaan Oromoo digits into text. Features a custom DeeperCNN architecture, SpecAugment for noise resilience, MLflow experiment tracking, and an interactive Streamlit UI for live inference. # 🎙️ SagalNet: Afaan Oromoo Spoken Digit Recognition **Real-time Spoken Digit Recognition using Deep Convolutional Neural Networks (CNNs).** 📚 Read the Docs | 🚀 Quick Start | 📊 Experiments --- ## 📖 Overview **SagalNet** implements a robust machine learning pipeline to recognize spoken digits (0-9) in **Afaan Oromoo**. It leverages modern deep learning techniques, including **Mel-Spectrograms** for feature extraction and a custom **DeeperCNN** architecture for high-accuracy classification. We focus on a complete **MLOps lifecycle**: - **Modular Codebase**: Clean separation of Data (`src/data`), Modeling (`src/models`), and UI. - **Experiment Tracking**: All runs are logged with **MLflow** (Metrics, Parameters, Models). - **Interactive UI**: A **Streamlit** app for real-time testing via microphone or file upload. ## ✨ Key Features - **🎙️ Live Recording**: Test the model instantly using your microphone. - **🧠 Advanced Architecture**: Custom `DeeperCNN` with BatchNorm, Dropout, and Adaptive Pooling. - **📈 SpecAugment**: Implements Time and Frequency masking for robust training. - **📊 Visualizations**: Real-time Mel-Spectrograms and Prediction Confidence bars. - **🛠️ Reproducible**: Full environment setup with `requirements.txt` and `venv`. ## 🚀 Quick Start ### 1. Clone & Setup ```bash git clone github.com cd SagalNet python -m venv venv source venv/bin/activate pip install -r requirements.txt ``` ### 2. Run the App Launch the interactive UI to test the model: ```bash streamlit run app.py ``` *Open localhost in your browser.* ### 3. Train the Model Train a new model from scratch: ```bash # Basic Training python run.py train --epochs 30 --model_type deeper # View Experiments mlflow ui ``` ## 🏗️ Architecture The system converts raw audio into visual representations (Mel-Spectrograms) which are then processed by a Deep CNN. ```mermaid graph LR A["🎙️ Audio Input"] --> B["🌊 Waveform"] B …