How allen23antony/audioshield-forensics works
AudioShield is an audio deepfake detection and forensic application. A browser or API client submits audio; the pipeline runs three model predictions plus metadata, signal, and supplementary segment analysis, then fuses findings and produces JSON artifacts and a PDF. The graph also preserves the AASIST research/training layer and the signal-analysis substeps. Relationships reflect the README and shown imports/calls; several subsystems are represented by their exact directory because their implementation was not sampled.
Main entry points: pipeline.py Analysis pipeline · predict.py Ensemble predictions · extractor.py Metadata extraction · signal.py Signal analysis
Components 24
Application flow
- Upload API
src/api/main.py - Analysis pipeline
src/inference/pipeline.py
Model inference
- Audio preprocessing
src/audio_preprocessing.py - Feature extraction
src/feature_extraction.py - Ensemble predictions
src/inference/predict.py - CNN training/model
src/train_cnn_final.py
Forensic analysis
- Metadata extraction
src/metadata/extractor.py - Metadata checks
src/metadata/analyzer.py - Signal analysis
src/forensics/signal.py - Temporal analysis
src/forensics/temporal.py - Signal visualizations
src/forensics/visualize.py - Segment scoring
src/forensics/segment.py - Integrity hashes
src/utils/hashing.py
Evidence and reporting
- Evidence fusion
src/fusion/combine.py - Fusion reports
src/fusion/report.py - Prediction artifacts
src/inference/save.py - Forensic PDF
src/report/generator.py
AASIST research
- AASIST model
src/aasist/model.py - AASIST dataset
src/aasist/dataset.py - AASIST configuration
src/aasist/config.py - AASIST training
src/aasist/train_balanced.py
Other components
- Browser user
- FFmpeg
- FFprobe / ExifTool
Connections 25
- Browser user → Upload API: uploads audio
README.md - Upload API → Analysis pipeline: invokes
src/api/main.py - Analysis pipeline → Ensemble predictions: requests predictions
src/inference/pipeline.py - Ensemble predictions → Audio preprocessing: preprocesses audio
src/inference/predict.py - Ensemble predictions → Feature extraction: extracts features
src/inference/predict.py - Ensemble predictions → CNN training/model: loads CNN
src/inference/predict.py - Ensemble predictions → AASIST model: runs AASIST
src/inference/predict.py - Analysis pipeline → Prediction artifacts: saves predictions
src/inference/pipeline.py - Analysis pipeline → Metadata extraction: extracts metadata
src/inference/pipeline.py - Analysis pipeline → Metadata checks: checks metadata
src/inference/pipeline.py - Metadata extraction → Integrity hashes: computes hashes
src/metadata/extractor.py - Metadata extraction → FFprobe / ExifTool: queries metadata
README.md - Analysis pipeline → Signal analysis: analyzes signal
src/inference/pipeline.py - Analysis pipeline → Segment scoring: scores segments
src/inference/pipeline.py - Segment scoring → Audio preprocessing: preprocesses audio
src/forensics/segment.py - Segment scoring → Ensemble predictions: reuses model inference
src/forensics/segment.py - Signal analysis → Temporal analysis: uses temporal checks
src/forensics/analyze.py - Signal analysis → Signal visualizations: produces plots
src/forensics/analyze.py - Analysis pipeline → Evidence fusion: combines evidence
src/inference/pipeline.py - Analysis pipeline → Fusion reports: writes fusion report
src/inference/pipeline.py - Analysis pipeline → Forensic PDF: generates PDF
src/inference/pipeline.py - Audio preprocessing → FFmpeg: decodes audio
README.md - AASIST training → AASIST dataset: loads examples
src/aasist/train_balanced.py - AASIST training → AASIST model: trains model
src/aasist/train_balanced.py - AASIST dataset → AASIST configuration: uses split settings
src/aasist/dataset.py