Voice Emotion Recognition

This project implements a machine learning pipeline for the classification of emotions from human voice recordings, using MFCC features as input.

It supports multiple supervised models (SVM, Logistic Regression), provides feature selection via Mutual Information, and includes AutoML experiments with TPOT and H2O. The codebase is modular, script-based (no notebooks), and designed for reproducibility.

Features

• MFCC feature extraction with Librosa
• Feature selection using Mutual Information
• Hyperparameter tuning via grid search and cross-validation
• AutoML experiments using TPOT and H2O
• Evaluation metrics: precision, recall, F1-score, accuracy
• Clean train/test split and preprocessing pipeline
• Structured, modular codebase without reliance on notebooks

Project Structure

voice-emotion-recognition/
├── data/
│   ├── raw/                ← Place your audio dataset here (ignored by Git)
├── models/                 ← Trained models are saved here
├── scripts/                ← All run scripts (training, grid search, etc.)
├── src/                    ← Core logic (feature extraction, preprocessing...)
│   ├── features.py
│   ├── preprocessing.py
│   ├── automl.py
│   ├── evaluate.py
│   └── utils.py
├── README.md               ← This file
├── requirements.txt        ← Python dependencies
└── .gitignore              ← Keeps datasets and model files out of Git

Installation

git clone https://github.com/Paul92150/voice-emotion-recognition.git
cd voice-emotion-recognition
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

Data Setup

The dataset should be placed in:

data/raw/wav/

Supported datasets include:

• Emo-DB (Berlin Database of Emotional Speech)
• CREMA-D
• RAVDESS
• Other compatible WAV-based datasets

The data/raw/wav/ folder is excluded from version control via .gitignore to prevent uploading personal or large files.

Emotion Label Mapping

By default, the system expects emotion labels to be encoded in the filenames (as is common with datasets like Emo-DB or CREMA-D). For example, using Emo-DB:

emotion_mapping = {
    'W': 'anger',
    'L': 'boredom',
    'E': 'disgust',
    'A': 'anxiety',
    'F': 'happiness',
    'T': 'sadness',
    'N': 'neutral'
}

If you are using a different dataset, make sure to:

• Update the emotion_mapping dictionary
• Adjust the label_position (character index in the filename)

These values are typically configured at the top of the relevant scripts (e.g., train_svm.py, automl_tpot.py).

Files with missing or unmapped labels are automatically skipped during training.

How to Run

Train an SVM Classifier

python scripts/train_svm.py

Run Grid Search for Optimal SVM Hyperparameters

python scripts/grid_search.py

Run TPOT AutoML

python scripts/automl_tpot.py

Perform Feature Selection (Mutual Information)

python scripts/feature_selection_mi.py

Example Results (TPOT AutoML)

Accuracy: 73%
F1-score (macro avg): 0.69
Best pipeline: MinMaxScaler → RFE → XGBoostClassifier

Note: These results were obtained on a relatively small dataset and are comparable to those of manually tuned SVMs.

Author

Paul Lemaire
MSc Candidate, CentraleSupélec
LinkedIn: https://www.linkedin.com/in/paul-lemaire-aa0369289
GitHub: https://github.com/Paul92150

Future Improvements

• Add support for CSV-based emotion labeling metadata
• Integrate ROC curve visualizations for evaluation
• Add support for additional datasets (e.g., TESS, SAVEE)
• Implement a configuration file (config.yaml) for parameter control
• Develop a test suite to validate pipeline integrity

License

This project is licensed under the MIT License. See the LICENSE file for details.