How ansh2709/emotion-classification works
Project type and purpose
• This is a full-stack machine-learning application for speech emotion recognition.
• It provides:
• Jupyter notebooks for training (DNN, XGBoost, RF)
• A CLI testing script (test_model.py)
• A Streamlit web application (app.py)
• Stored model artifacts and configuration under trained_models/
File structure
Top level:
• classification-files/
• complete ipynb(AdvancedDNN).ipynb
• complete_ipynb(Xgboost).ipynb
• app.py (Streamlit UI)
• test_model.py (CLI interface)
• requirements.txt
• trained models/ (emotion_model.h5, scaler, encoder, config, etc.)
• demo/ (demo_video.mp4)
• pretrained-concepts-ipynb/ (reference notebooks)
Key observations:
• Separation of concerns: training (notebooks), deployment (app.py, test_model.py), artifacts (trained models)
• No separate “backend” HTTP service—UI and CLI directly load artifacts and run inference
• Configuration files and pickled objects under trained models indicate a persistence layer
README for additional insights
• The README outlines a linear “Raw Audio → Preprocessing → Feature Extraction → Model Training → Deployment” flow.
• Technologies: Python 3.8+, TensorFlow 2.x, XGBoost, scikit-learn, Streamlit.
• Feature extraction: MFCC, Chroma, Spectral Centroid, etc.
• Two deployment targets: CLI and Streamlit web UI.
Detailed capture of essential architectural elements
• Don’t just group everything under “backend.” Split preprocessing vs feature extraction vs inference.
• That both Run paths (CLI and Web) share the same core pipeline modules.
• Call out external dependencies (TensorFlow, XGBoost, Streamlit) next to each component.
• Configuration files (config.pkl, scaler.pkl, encoder.pkl) as part of the persistence layer.