W2V2-AASIST

A wav2vec 2.0 (XLS-R 300M) + AASIST anti-spoofing model, from "Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation" (Tak, Todisco, Wang, Jung, Yamagishi & Evans, Odyssey 2022). A self-supervised XLS-R front-end is fine-tuned end-to-end with an AASIST spectro-temporal graph-attention back-end. The model takes a raw speech waveform and returns a score where higher = more bona fide.

• Code: https://github.com/TakHemlata/SSL_Anti-spoofing
• Paper: https://arxiv.org/abs/2202.12233
• Parameters: 317,837,800 (317.84 M)
• Checkpoint: LA_model.pth (the LA variant)

The exact wrapper used to produce the Arena scores is in w2v2_aasist.py; the network definition is in _net.py.

Architecture

1. wav2vec 2.0 XLS-R (300M) front-end — a self-supervised transformer (fairseq Wav2Vec2Model) producing 1024-d frame features, fine-tuned end-to-end with the rest of the network.
2. AASIST back-end — the XLS-R features are projected to 128-d, max-pooled, passed through a RawNet2-style residual encoder, then heterogeneous stacking graph-attention layers (HS-GAL) over spectral and temporal sub-graphs with a learnable master node and graph pooling.
3. The 2-logit output is read at index 1 = bona fide.

How it was trained

• Data: ASVspoof 2019 Logical Access (LA), with RawBoost data augmentation.
• Input length: raw audio at 16 kHz cropped/padded to 64,600 samples (~4.04 s).
• Output: 2-class logits; the bona-fide logit (index 1) is the score.

See the source repository for the full training and evaluation code.

Benchmark result (Speech Anti-Spoofing Arena)

Evaluated through the reproducible Speech Anti-Spoofing Arena. Scores were computed with a deterministic first-64,600-sample window (no random crop), so the numbers are exactly reproducible from the pinned score file.

Dataset	Split	EER %	Trials	Skipped	Notes
ASVspoof2019_LA	test	0.22	71,237	0	in-domain (training data)
ASVspoof2021_LA	test	8.11	181,566	0	cross-dataset generalization
ASVspoof2021_DF	test	8.32	611,829	0	cross-dataset generalization
InTheWild	test	11.22	31,779	0	out-of-domain (real-world deepfakes)
CD-ADD	test	38.57	20,786	0	out-of-domain (modern neural-TTS)
SONAR	test	46.12	3,948	0	out-of-domain (diverse deepfake sources)
LibriSeVoc	test	11.21	18,487	0	out-of-domain (LibriTTS neural vocoders)
CFAD	test	17.28	62,999	0	out-of-domain (Chinese fake-audio detection)
CVoiceFake_small	test	21.79	138,136	0	out-of-domain (multilingual vocoded TTS)
ASVspoof5	test	16.25	680,774	0	out-of-domain (crowdsourced TTS/VC + adversarial)

The self-supervised XLS-R front-end generalizes markedly better to unseen attacks than raw-waveform baselines — most strikingly on InTheWild (11.22 %) and CD-ADD (38.57 %), where lightweight CNN models degrade much further.

Usage

The checkpoint is a state_dict for the Model network defined in _net.py. Constructing the network requires the base XLS-R 300M checkpoint xlsr2_300m.pt next to the wrapper (only used to build the wav2vec 2.0 architecture; every weight is then overwritten by LA_model.pth):

wget https://dl.fbaipublicfiles.com/fairseq/wav2vec/xlsr2_300m.pt

The input must be exactly 64,600 samples at 16 kHz mono — window the waveform with pad_fixed (first 64,600 samples, tile-repeat if shorter).

import numpy as np
from w2v2_aasist import W2V2AASIST   # _net.py + w2v2_aasist.py are in this repo

m = W2V2AASIST()
m.load()                                          # loads LA_model.pth (+ xlsr2_300m.pt)
audio = np.random.randn(48000).astype(np.float32) # float32 mono 16 kHz
print(m.score_batch([audio], [16000])[0])         # higher = more bona fide
m.unload()

Internally the wrapper windows the input, runs the network, and returns logits[:, 1] (class 1 = bona fide). w2v2_aasist.py is the exact speech_spoof_bench model that produced the Arena scores.txt.

Citation

This model / paper:

@inproceedings{tak2022automatic,
  title={Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation},
  author={Tak, Hemlata and Todisco, Massimiliano and Wang, Xin and Jung, Jee-weon and Yamagishi, Junichi and Evans, Nicholas},
  booktitle={The Speaker and Language Recognition Workshop (Odyssey 2022)},
  pages={112--119},
  year={2022}
}

AASIST back-end:

@inproceedings{jung2022aasist,
  title={{AASIST}: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks},
  author={Jung, Jee-weon and Heo, Hee-Soo and Tak, Hemlata and Shim, Hye-jin and Chung, Joon Son and Lee, Bong-Jin and Yu, Ha-Jin and Evans, Nicholas},
  booktitle={ICASSP 2022},
  pages={6367--6371},
  year={2022},
  organization={IEEE}
}

License

MIT — see the source repository.