🎙️ 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

git clone https://github.com/abdulmunimjemal/SagalNet.git
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:

streamlit run app.py

Open http://localhost:8501 in your browser.

3. Train the Model

Train a new model from scratch:

# 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.

graph LR
    A["🎙️ Audio Input"] --> B["🌊 Waveform"]
    B --> C["🖼️ Mel-Spectrogram"]
    C --> D["🧠 DeeperCNN"]
    D --> E["📊 Probability Distribution"]
    E --> F["✅ Prediction"]

Loading

See docs/02_architecture.md for detailed diagrams.

📊 Results

Model	Epochs	Accuracy	F1-Score
SimpleCNN	30	~87%	0.87
DeeperCNN	50	91.94%	0.9194

📂 Project Structure

├── app.py                  # Streamlit UI Entry point
├── notebooks/              # Jupyter Notebooks for analysis
├── src/
│   ├── data/               # Dataset loading & augmentation
│   ├── models/             # CNN Architectures & Training Loop
│   └── utils/              # Helper scripts
├── docs/                   # Detailed Documentation
├── requirements.txt        # Dependencies
└── run.py                  # CLI Entry point