Models

Jabberjay bundles ten model families. Each is downloaded from Hugging Face Hub on first use and cached locally.

Choosing a model

Model	Type	Datasets	Requires
VIT	Vision Transformer	ASVspoof2019, ASVspoof5, VoxCelebSpoof	`dataset`, `visualisation`
AST	Audio Spectrogram Transformer	ASVspoof2019, ASVspoof5, VoxCelebSpoof	`dataset`
Spectra0	Wav2Vec2 + ECAPA-TDNN	ASVspoof19/21, ASVspoof5, In-the-Wild	—
SpectraAASIST	Wav2Vec2 + AASIST	ASVspoof19/21, ASVspoof5, In-the-Wild	—
SpectraAASIST3	Wav2Vec2 + KAN-AASIST	ASVspoof19/21, ASVspoof5, In-the-Wild	—
Wav2Vec2	Self-supervised transformer	ASVspoof2019	—
HuBERT	Self-supervised transformer	In-The-Wild	—
WavLM	Self-supervised transformer	Mixed deepfake	—
RawNet2	End-to-end CNN	ASVspoof 2021	—
Classical	KNN classifier	ASVspoof2019	—

Simple rule of thumb:

For the lowest error rate on In-the-Wild audio — use SpectraAASIST3 (EER 0.961%)
For a quick, general-purpose result — use WavLM or HuBERT
For a lightweight baseline with no deep learning — use Classical
To sweep all models and compare — use jj.load() once, then call jj.detect() for each

VIT

Vision Transformer classifiers that convert the audio to a 2D image (spectrogram) and classify it visually.

Nine variants, covering three visualisation types × three training datasets:

Visualisation	ASVspoof2019	ASVspoof5	VoxCelebSpoof
ConstantQ	✓	✓	✓
MelSpectrogram	✓	✓	✓
MFCC	✓	✓	✓

jj.detect("audio.wav", model="VIT", dataset="VoxCelebSpoof", visualisation="ConstantQ")
jj.detect("audio.wav", model="VIT", dataset="ASVspoof2019", visualisation="MFCC")
jj.detect("audio.wav", model="VIT", dataset="ASVspoof5", visualisation="MelSpectrogram")

Visualisations:

ConstantQ — Constant-Q transform; good frequency resolution across the full spectrum
MelSpectrogram — Mel-scaled spectrogram; perceptually motivated, widely used in speech
MFCC — Mel-frequency cepstral coefficients; compact and speech-focused

AST

Audio Spectrogram Transformer. Applies a transformer directly to a patch-based spectrogram without a CNN backbone.

Available for ASVspoof2019, ASVspoof5, and VoxCelebSpoof datasets.

jj.detect("audio.wav", model="AST")  # defaults to VoxCelebSpoof
jj.detect("audio.wav", model="AST", dataset="VoxCelebSpoof")
jj.detect("audio.wav", model="AST", dataset="ASVspoof2019")

Note

dataset is optional for AST and defaults to VoxCelebSpoof if not specified.

Spectra0

lab260/spectra_0

SSL encoder (Wav2Vec2-XLS-R-300M) → MLP projection → ECAPA-TDNN binary classifier, trained across multiple anti-spoofing benchmarks.

Benchmark	EER (%)
ASVspoof19 LA	0.181
ASVspoof21 LA	6.475
ASVspoof21 DF	5.410
ASVspoof5	14.426
ADD2022	14.716
In-the-Wild	1.026

jj.detect("audio.wav", model="Spectra0")

Note

On first run, Spectra0 downloads both its own weights and the facebook/wav2vec2-xls-r-300m SSL encoder (~1.2 GB total). Both are cached by HuggingFace Hub after the first download.

SpectraAASIST

lab260/Spectra-AASIST

SSL encoder (Wav2Vec2-XLS-R-300M) → MLP projection → AASIST graph attention network with standard linear layers.

Benchmark	EER (%)
ASVspoof19 LA	0.159
ASVspoof21 LA	5.164
ASVspoof21 DF	2.568
ASVspoof5	14.056
ADD2022	15.205
In-the-Wild	1.461

jj.detect("audio.wav", model="SpectraAASIST")

Note

Downloads facebook/wav2vec2-xls-r-300m on first run (~1.2 GB total, cached by HuggingFace Hub).

SpectraAASIST3

lab260/Spectra-AASIST3

SSL encoder (Wav2Vec2-XLS-R-300M) → MLP projection → AASIST graph attention network with KAN (Kolmogorov-Arnold Network) linear layers — the strongest model in the Spectra family on In-the-Wild audio.

Benchmark	EER (%)
ASVspoof19 LA	0.723
ASVspoof21 LA	4.506
ASVspoof21 DF	1.998
ASVspoof5	13.820
ADD2022	15.187
In-the-Wild	0.961

jj.detect("audio.wav", model="SpectraAASIST3")

Note

Downloads facebook/wav2vec2-xls-r-300m on first run (~1.2 GB total, cached by HuggingFace Hub).

Wav2Vec2

Gustking/wav2vec2-large-xlsr-deepfake-audio-classification

Wav2Vec2-XLSR-300M fine-tuned on ASVspoof2019. EER 4.01%.

jj.detect("audio.wav", model="Wav2Vec2")

HuBERT

abhishtagatya/hubert-base-960h-itw-deepfake

HuBERT-base fine-tuned on the In-The-Wild dataset. EER 1.43%.

jj.detect("audio.wav", model="HuBERT")

WavLM

DavidCombei/wavLM-base-Deepfake_V2

WavLM-base fine-tuned on a mixed deepfake dataset.

jj.detect("audio.wav", model="WavLM")

RawNet2

End-to-end anti-spoofing network via rawnet2-antispoofing, with weights from ASVspoof 2021. Operates directly on raw waveforms — no feature extraction step.

jj.detect("audio.wav", model="RawNet2")

Note

scores is None for RawNet2 — only label, is_bonafide, and confidence are populated.

Classical

Feature-based KNN classifier trained on ASVspoof2019. Extracts hand-crafted audio features (MFCCs, spectral features) and classifies with k-nearest neighbours. Fast and dependency-light compared to the transformer models.

jj.detect("audio.wav", model="Classical")

Note

scores is None for Classical — only label, is_bonafide, and confidence are populated.