You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

AugAll Speech Model

Speech deepfake / spoof detector trained with full augmentation (AugAll) on universal speech data.

Item	Value
Checkpoint	`checkpoint_6_augall.pth` (epoch 6)
Frontend	Wav2Vec2 XLS-R-2B (HF)
Backend	AASIST
Loss	AMSoftmax (2-class)
Input	16 kHz mono, padded/truncated to 64600 samples (~4 s)
Labels	`bonafide` (real) / `spoof` (fake)

Requires the DeepFense framework (model registry + AASIST backend):

git clone <DeepFense-repo> /path/to/DeepFense
export DEEPFENSE_ROOT=/path/to/DeepFense
pip install -r requirements.txt
pip install -e $DEEPFENSE_ROOT

Accept the XLS-R-2B license and login:

huggingface-cli login

python infer.py path/to/audio.wav
python infer.py path/to/audio.wav --device cuda

Example output:

bonafide (bonafide=0.9823, spoof=0.0177)

checkpoint_6_augall.pth — full model weights (~25 GB, includes finetuned XLS-R-2B + AASIST head)
config.yaml — model architecture config
infer.py — single-file inference script

Trained with concatenated augmentation pipeline (rawboost, RIR, noise, speed perturb)
Universal speech deepfake parquet mixtures under /ds-slt/audio/universal_speech_df_data/