How sampada-11/voice-deepfake-detector works
Voice Deepfake Detector is a local Flask application backed by a classical acoustic-feature/SVM pipeline. A user uploads audio in the browser; Flask validates it, extracts 36 features, predicts REAL or FAKE with a confidence score, logs the result, and returns it for display. The repository also includes dataset inspection and construction, baseline and conversion-aware experiments, and final-model training. Source excerpts confirm the main inference and data-building calls; broader ML-stage links follow the README, while unsampled implementation details are not assumed.
Main entry points: app.py Flask application · data Feature dataset · build_dataset.py Dataset builder · train_final_model.py Final model training
Components 18
Web detection
- Browser UI
static/js/main.js - Web pages
templates - Flask application
app.py - Prediction response
Acoustic features
- Feature extraction
ml/data_processing/feature_extraction.py - 36-feature vector
Dataset preparation
- Dataset inspection
ml/data_processing/inspect_dataset.py - Dataset builder
ml/data_processing/build_dataset.py - Feature dataset
data
Modeling and evaluation
- Baseline experiments
ml/train.py - Final model training
ml/train_final_model.py - Conversion-aware evaluation
ml/evaluation/conversion_aware_evaluation.py - Saved SVM pipeline
models/detector.pkl - Feature metadata
models/model_metadata.json
Runtime state
- Detection history
Other components
- User
- Uploaded audio
- DeepVoice dataset
Connections 19
- User → Browser UI: selects audio
README.md - Browser UI → Flask application: submits audio
README.md - Flask application → Feature extraction: extracts features
app.py - Flask application → Saved SVM pipeline: loads and predicts
app.py - Feature extraction → 36-feature vector: returns vector
README.md - Flask application → Detection history: writes and reads
app.py - Flask application → Prediction response: returns prediction
README.md - Prediction response → Browser UI: displays result
README.md - User → Web pages: views pages
README.md - DeepVoice dataset → Dataset inspection: inspects samples
ml/data_processing/inspect_dataset.py - DeepVoice dataset → Dataset builder: builds from samples
ml/data_processing/build_dataset.py - Dataset builder → Feature extraction: extracts features
ml/data_processing/build_dataset.py - Dataset builder → Feature dataset: writes feature data
ml/data_processing/build_dataset.py - Feature dataset → Baseline experiments: feeds experiments
ml/train.py - Feature dataset → Final model training: feeds training
ml/train_final_model.py - Feature dataset → Conversion-aware evaluation: feeds evaluation
README.md - Final model training → Saved SVM pipeline: saves pipeline
README.md - Final model training → Feature metadata: saves feature order
README.md - Baseline experiments → Conversion-aware evaluation: compares generalization
README.md