How clovaai/aasist works
AASIST is a Python framework for training and evaluating audio anti-spoofing models on ASVspoof data. The main workflow loads experiment configuration and audio trials, selects a model architecture, then trains or runs evaluation, producing countermeasure scores and EER/t-DCF metrics. The README also documents dataset downloading, pretrained-model evaluation, and four model choices. Source excerpts show the main runtime wiring; some training details and configuration-to-weight mappings are documented but not fully sampled.
Main entry points: main.py Experiment runner · config Experiment configs · download_dataset.py Dataset download · data_utils.py Trial loading
Components 14
Experiment workflow
- Experiment runner
main.py - Experiment configs
config - Score output
main.py
Data preparation
- ASVspoof audio
- Dataset download
download_dataset.py - Trial loading
data_utils.py
Detection models
- AASIST models
models/AASIST.py - RawNet2 baseline
models/RawNet2Spoof.py - RawGAT-ST baseline
models/RawNetGatSpoofST.py - Pretrained weights
models/weights
Optimization and metrics
- Training utilities
utils.py - Evaluation metrics
evaluation.py - ASV reference scores
Other components
- Researcher
Connections 15
- Researcher → Experiment runner: runs
README.md - Researcher → Dataset download: requests download
README.md - Dataset download → ASVspoof audio: downloads archive
download_dataset.py - Researcher → Experiment configs: selects config
README.md - Experiment configs → Experiment runner: configures run
README.md - ASVspoof audio → Trial loading: provides audio
data_utils.py - Experiment runner → Trial loading: builds loaders
main.py - Experiment runner → AASIST models: loads architecture
main.py - Experiment runner → RawNet2 baseline: loads architecture
- Experiment runner → RawGAT-ST baseline: loads architecture
- Experiment runner → Pretrained weights: loads pretrained weights
README.md - Experiment runner → Training utilities: uses training utilities
main.py - Experiment runner → Score output: writes scores
main.py - Experiment runner → Evaluation metrics: computes metrics
main.py - ASV reference scores → Evaluation metrics: supplies ASV scores