AASIST-L is the
lightweight variant of AASIST audio anti-spoofing
(voice-deepfake detection) from
"AASIST: Audio Anti-Spoofing using Integrated
Spectro-Temporal Graph Attention Networks" (Jung et al., ICASSP 2022). It uses
the upstream
clovaai/aasist ASVspoof2019 LA
pretrained
AASIST-L checkpoint. The model takes a raw speech waveform and
returns a score where
higher = more bona fide.
This repo is self-contained for inference: the network definition is in
_net.py (identical to the full AASIST) and the exact wrapper used
to produce the Arena scores in
aasist_l.py. AASIST-L shares
the AASIST architecture but with a narrower residual stack and graph dimensions
(~85k params vs ~298k).
AASIST operates directly on the raw waveform: a sinc-convolution front-end and a
RawNet2-style residual encoder produce a spectro-temporal feature map, which is
modelled by heterogeneous stacking graph attention layers over spectral and
temporal sub-graphs with a learnable max/average readout, followed by a 2-class
output (bona fide vs. spoof). The Arena score is the bona-fide logit. The "-L"
variant narrows the residual channels (…[32,24],[24,24]) and graph dims
([24,32]).
Inference uses a deterministic first-64600-sample window (no random crop),
matching the upstream data_utils.pad() used at eval. Audio is provided as
float32 mono at 16 kHz (no resampling in the wrapper).
1from aasist_l import AASIST_L
2m = AASIST_L(); m.load()
3scores = m.score_batch([wav], [16000]) # higher = more bona fide
The in-domain ASVspoof2019 LA result (~0.99%) reproduces the paper's reported
AASIST-L EER. AASIST-L matches the full AASIST closely at ~3.5× fewer parameters.
MIT (inherited from clovaai/aasist; see
LICENSE).
Maintained by Kirill Borodin (SpeechAntiSpoofingBenchmarks).
