Tidy submissions: ★ inference (maintainers re-ran these models), one params value per model, drop unverified duplicate files

submissions/aasist-l.yaml

@@ -2,23 +2,26 @@ schema_version: 4
 system:
   name: AASIST-L
   slug: aasist-l
-  description: 'AASIST-L: the lightweight variant of AASIST (audio anti-spoofing using integrated spectro-temporal
-    graph attention networks). Same architecture as AASIST — sinc-convolution front-end, RawNet2-style
-    residual encoder, and heterogeneous stacking graph attention over spectral and temporal sub-graphs
-    with a learnable readout — but with a narrower residual stack and graph dimensions (~85k params vs
-    AASIST''s ~298k). Official clovaai/aasist ASVspoof2019 LA pretrained checkpoint, FP32, deterministic
-    first-64600-sample window (no random crop).'
+  description: 'AASIST-L: the lightweight variant of AASIST (audio anti-spoofing using
+    integrated spectro-temporal graph attention networks). Same architecture as AASIST
+    — sinc-convolution front-end, RawNet2-style residual encoder, and heterogeneous
+    stacking graph attention over spectral and temporal sub-graphs with a learnable
+    readout — but with a narrower residual stack and graph dimensions (~85k params
+    vs AASIST''s ~298k). Official clovaai/aasist ASVspoof2019 LA pretrained checkpoint,
+    FP32, deterministic first-64600-sample window (no random crop).'
   code: https://github.com/clovaai/aasist
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/AASIST-L/blob/e4185b270ec20077c918e06a45093717a1bd5e30/AASIST-L.pth
   params_millions: 0.085306
   paper:
     arxiv_id: '2110.01200'
     url: https://arxiv.org/abs/2110.01200
-    bibtex: "@inproceedings{jung2022aasist,\n  title={{AASIST}: Audio Anti-Spoofing Using Integrated Spectro-Temporal\
-      \ Graph Attention Networks},\n  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},\n  booktitle={ICASSP\
-      \ 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},\n\
-      \  pages={6367--6371},\n  year={2022},\n  organization={IEEE}\n}\n"
+    bibtex: "@inproceedings{jung2022aasist,\n  title={{AASIST}: Audio Anti-Spoofing\
+      \ Using Integrated Spectro-Temporal Graph Attention Networks},\n  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},\n  booktitle={ICASSP\
+      \ 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal\
+      \ Processing (ICASSP)},\n  pages={6367--6371},\n  year={2022},\n  organization={IEEE}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -35,12 +38,13 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: Lightweight AASIST-L variant (not the full AASIST). Same _net.py Model class as AASIST; differs
-  only in model_config (filts [..,[32,24],[24,24]], gat_dims [24,32], pool_ratios [0.4,0.5,0.7,0.5]) and
-  checkpoint. Deterministic first-64600-sample window (no random crop), matching clovaai/aasist data_utils.pad()
-  used at eval. Checkpoint mirrored to SpeechAntiSpoofingBenchmarks/AASIST-L (pinned at publish).
+notes: Lightweight AASIST-L variant (not the full AASIST). Same _net.py Model class
+  as AASIST; differs only in model_config (filts [..,[32,24],[24,24]], gat_dims [24,32],
+  pool_ratios [0.4,0.5,0.7,0.5]) and checkpoint. Deterministic first-64600-sample
+  window (no random crop), matching clovaai/aasist data_utils.pad() used at eval.
+  Checkpoint mirrored to SpeechAntiSpoofingBenchmarks/AASIST-L (pinned at publish).

submissions/aasist.yaml

@@ -2,21 +2,24 @@ schema_version: 4
 system:
   name: AASIST
   slug: aasist
-  description: 'AASIST: audio anti-spoofing using integrated spectro-temporal graph attention networks.
-    Sinc-convolution front-end, RawNet2-style residual encoder, and heterogeneous stacking graph attention
-    over spectral and temporal sub-graphs with a learnable readout. Official clovaai/aasist ASVspoof2019
-    LA pretrained checkpoint, FP32, deterministic first-64600-sample window (no random crop).'
+  description: 'AASIST: audio anti-spoofing using integrated spectro-temporal graph
+    attention networks. Sinc-convolution front-end, RawNet2-style residual encoder,
+    and heterogeneous stacking graph attention over spectral and temporal sub-graphs
+    with a learnable readout. Official clovaai/aasist ASVspoof2019 LA pretrained checkpoint,
+    FP32, deterministic first-64600-sample window (no random crop).'
   code: https://github.com/clovaai/aasist
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/AASIST/blob/e842653505c2832ac9f46bbf56173b0f54ef82a7/AASIST.pth
   params_millions: 0.297866
   paper:
     arxiv_id: '2110.01200'
     url: https://arxiv.org/abs/2110.01200
-    bibtex: "@inproceedings{jung2022aasist,\n  title={{AASIST}: Audio Anti-Spoofing Using Integrated Spectro-Temporal\
-      \ Graph Attention Networks},\n  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},\n  booktitle={ICASSP\
-      \ 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},\n\
-      \  pages={6367--6371},\n  year={2022},\n  organization={IEEE}\n}\n"
+    bibtex: "@inproceedings{jung2022aasist,\n  title={{AASIST}: Audio Anti-Spoofing\
+      \ Using Integrated Spectro-Temporal Graph Attention Networks},\n  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},\n  booktitle={ICASSP\
+      \ 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal\
+      \ Processing (ICASSP)},\n  pages={6367--6371},\n  year={2022},\n  organization={IEEE}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -33,11 +36,11 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: Official AASIST variant only (not AASIST-L). Deterministic first-64600-sample window (no random
-  crop), matching clovaai/aasist data_utils.pad() used at eval. Checkpoint mirrored to SpeechAntiSpoofingBenchmarks/AASIST
-  (pinned).
+notes: Official AASIST variant only (not AASIST-L). Deterministic first-64600-sample
+  window (no random crop), matching clovaai/aasist data_utils.pad() used at eval.
+  Checkpoint mirrored to SpeechAntiSpoofingBenchmarks/AASIST (pinned).

submissions/aasist3.yaml

@@ -2,24 +2,27 @@ schema_version: 4
 system:
   name: AASIST3
   slug: aasist3
-  description: 'KAN-enhanced AASIST speech deepfake detector with a wav2vec 2.0 (XLS-R-53) self-supervised
-    front-end. SSL features are projected to 128-d through a Kolmogorov-Arnold (KAN) bridge, processed
-    by a RawNet2-style residual encoder and spectro-temporal graph-attention layers, and classified by
-    a four-branch inference head with a KAN output layer. ASVspoof 2024 Challenge system; the published
-    lab260/AASIST3 checkpoint (note: these weights differ from the paper results). FP32, preemphasis (0.97),
-    deterministic first-64600-sample window (no random crop). score = output logit for class 1 (bona fide).'
+  description: 'KAN-enhanced AASIST speech deepfake detector with a wav2vec 2.0 (XLS-R-53)
+    self-supervised front-end. SSL features are projected to 128-d through a Kolmogorov-Arnold
+    (KAN) bridge, processed by a RawNet2-style residual encoder and spectro-temporal
+    graph-attention layers, and classified by a four-branch inference head with a
+    KAN output layer. ASVspoof 2024 Challenge system; the published lab260/AASIST3
+    checkpoint (note: these weights differ from the paper results). FP32, preemphasis
+    (0.97), deterministic first-64600-sample window (no random crop). score = output
+    logit for class 1 (bona fide).'
   code: https://github.com/mtuciru/AASIST3
   checkpoint: https://huggingface.co/lab260/AASIST3
   params_millions: 321.7495
   paper:
     arxiv_id: '2408.17352'
     url: https://arxiv.org/abs/2408.17352
-    bibtex: "@inproceedings{borodin24_asvspoof,\n  title={AASIST3: KAN-enhanced AASIST speech deepfake\
-      \ detection using SSL features and additional regularization for the ASVspoof 2024 Challenge},\n\
-      \  author={Borodin, Kirill and Kudryavtsev, Vasiliy and Korzh, Dmitrii and Efimenko, Alexey and\
-      \ Mkrtchian, Grach and Gorodnichev, Mikhail and Rogov, Oleg Y.},\n  booktitle={The Automatic Speaker\
-      \ Verification Spoofing Countermeasures Workshop (ASVspoof 2024)},\n  pages={48--55},\n  year={2024},\n\
-      \  doi={10.21437/ASVspoof.2024-8}\n}\n"
+    bibtex: "@inproceedings{borodin24_asvspoof,\n  title={AASIST3: KAN-enhanced AASIST\
+      \ speech deepfake detection using SSL features and additional regularization\
+      \ for the ASVspoof 2024 Challenge},\n  author={Borodin, Kirill and Kudryavtsev,\
+      \ Vasiliy and Korzh, Dmitrii and Efimenko, Alexey and Mkrtchian, Grach and Gorodnichev,\
+      \ Mikhail and Rogov, Oleg Y.},\n  booktitle={The Automatic Speaker Verification\
+      \ Spoofing Countermeasures Workshop (ASVspoof 2024)},\n  pages={48--55},\n \
+      \ year={2024},\n  doi={10.21437/ASVspoof.2024-8}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -36,14 +39,16 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: AASIST3 (KAN-enhanced AASIST + wav2vec 2.0 XLS-R-53 SSL front-end) from mtuciru/AASIST3, weights
-  from lab260/AASIST3 (model.safetensors, self-contained incl. the SSL encoder). Loaded via PyTorchModelHubMixin.from_pretrained;
-  the XLS-R architecture is built from the wav2vec2-large-xlsr-53 config then every weight is overwritten
-  by the checkpoint. Preemphasis (0.97) applied to the full waveform before a deterministic first-64600-sample
-  window, matching the source datasets/generic.py eval pipeline. score = logit for class 1 (bona fide);
-  higher = more bona fide (source label map bonafide=1, validation uses outputs[:, 1]).
+notes: AASIST3 (KAN-enhanced AASIST + wav2vec 2.0 XLS-R-53 SSL front-end) from mtuciru/AASIST3,
+  weights from lab260/AASIST3 (model.safetensors, self-contained incl. the SSL encoder).
+  Loaded via PyTorchModelHubMixin.from_pretrained; the XLS-R architecture is built
+  from the wav2vec2-large-xlsr-53 config then every weight is overwritten by the checkpoint.
+  Preemphasis (0.97) applied to the full waveform before a deterministic first-64600-sample
+  window, matching the source datasets/generic.py eval pipeline. score = logit for
+  class 1 (bona fide); higher = more bona fide (source label map bonafide=1, validation
+  uses outputs[:, 1]).

submissions/amsdf.yaml

@@ -2,10 +2,12 @@ schema_version: 4
 system:
   name: AMSDF
   slug: amsdf
-  description: 'AMSDF (Audio Multi-view Spoofing Detection Framework): captures intra- and inter-view spoofing cues from audio-emotion-text
-    correlations. XLS-R (wav2vec2-xls-r-300m) text view + ACRNN emotion view (logfbank+delta+delta2) + raw-audio graph view,
-    fused by heterogeneous graph attention. Raw-waveform (+ derived emotion spectrogram), FP32, first-64600-sample (4 s) window.
-    Checkpoint M1 (best-EER seed). Higher score = more bona fide.
+  description: 'AMSDF (Audio Multi-view Spoofing Detection Framework): captures intra-
+    and inter-view spoofing cues from audio-emotion-text correlations. XLS-R (wav2vec2-xls-r-300m)
+    text view + ACRNN emotion view (logfbank+delta+delta2) + raw-audio graph view,
+    fused by heterogeneous graph attention. Raw-waveform (+ derived emotion spectrogram),
+    FP32, first-64600-sample (4 s) window. Checkpoint M1 (best-EER seed). Higher score
+    = more bona fide.
 
     '
   code: https://github.com/ItzJuny/AMSDF
@@ -14,10 +16,12 @@ system:
   paper:
     arxiv_id: 10.1109/TIFS.2024.3431888
     url: https://ieeexplore.ieee.org/document/10605761
-    bibtex: "@article{wu2024audio,\n  author  = {Wu, Junyan and Yin, Qilin and Sheng, Ziqi and Lu, Wei and Huang, Jiwu and\
-      \ Li, Bin},\n  journal = {IEEE Transactions on Information Forensics and Security},\n  title   = {Audio Multi-view Spoofing\
-      \ Detection Framework Based on Audio-Text-Emotion Correlations},\n  year    = {2024},\n  volume  = {19},\n  pages  \
-      \ = {7133--7146},\n  doi     = {10.1109/TIFS.2024.3431888}\n}\n"
+    bibtex: "@article{wu2024audio,\n  author  = {Wu, Junyan and Yin, Qilin and Sheng,\
+      \ Ziqi and Lu, Wei and Huang, Jiwu and Li, Bin},\n  journal = {IEEE Transactions\
+      \ on Information Forensics and Security},\n  title   = {Audio Multi-view Spoofing\
+      \ Detection Framework Based on Audio-Text-Emotion Correlations},\n  year   \
+      \ = {2024},\n  volume  = {19},\n  pages   = {7133--7146},\n  doi     = {10.1109/TIFS.2024.3431888}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -34,10 +38,10 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-26'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-26'
-notes: AMSDF M1 checkpoint (best-EER seed). Multi-view XLS-R + ACRNN emotion + graph fusion. Score = P(real)-P(fake). First-64600-sample
-  window.
+notes: AMSDF M1 checkpoint (best-EER seed). Multi-view XLS-R + ACRNN emotion + graph
+  fusion. Score = P(real)-P(fake). First-64600-sample window.

submissions/clad.yaml

@@ -2,9 +2,11 @@ schema_version: 4
 system:
   name: CLAD
   slug: clad
-  description: 'CLAD (Contrastive Learning-based Audio Deepfake detector): AASIST encoder trained with contrastive learning
-    for robustness against manipulation attacks, plus a linear classifier over the 160-dim AASIST embedding. Raw-waveform,
-    FP32, deterministic first-64600-sample (4 s) window (no random crop). Higher score = more bona fide.
+  description: 'CLAD (Contrastive Learning-based Audio Deepfake detector): AASIST
+    encoder trained with contrastive learning for robustness against manipulation
+    attacks, plus a linear classifier over the 160-dim AASIST embedding. Raw-waveform,
+    FP32, deterministic first-64600-sample (4 s) window (no random crop). Higher score
+    = more bona fide.
 
     '
   code: https://github.com/CLAD23/CLAD
@@ -13,9 +15,11 @@ system:
   paper:
     arxiv_id: '2404.15854'
     url: https://arxiv.org/abs/2404.15854
-    bibtex: "@article{wu2024clad,\n  title   = {CLAD: Robust Audio Deepfake Detection Against Manipulation Attacks with Contrastive\
-      \ Learning},\n  author  = {Wu, Haolin and Chen, Jing and Du, Ruiying and Wu, Cong and He, Kun and Shang, Xingcan and\
-      \ Ren, Hao and Xu, Guowen},\n  journal = {arXiv preprint arXiv:2404.15854},\n  year    = {2024}\n}\n"
+    bibtex: "@article{wu2024clad,\n  title   = {CLAD: Robust Audio Deepfake Detection\
+      \ Against Manipulation Attacks with Contrastive Learning},\n  author  = {Wu,\
+      \ Haolin and Chen, Jing and Du, Ruiying and Wu, Cong and He, Kun and Shang,\
+      \ Xingcan and Ren, Hao and Xu, Guowen},\n  journal = {arXiv preprint arXiv:2404.15854},\n\
+      \  year    = {2024}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -32,9 +36,10 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-26'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-26'
-notes: Pretrained CLAD model CLAD_150_10_2310.pth.tar (AASIST encoder, contrastive). Deterministic first-64600-sample window.
+notes: Pretrained CLAD model CLAD_150_10_2310.pth.tar (AASIST encoder, contrastive).
+  Deterministic first-64600-sample window.

submissions/df-arena-1b.yaml

@@ -2,22 +2,25 @@ schema_version: 4
 system:
   name: DF Arena 1B
   slug: df-arena-1b
-  description: RAPTOR universal anti-spoofing model. A wav2vec 2.0 XLS-R 1B self-supervised front-end
-    whose per-layer hidden states are combined by learnable attention pooling (a layer-wise sigmoid gate
-    over an attention-pooled summary), then passed through a 4-block Conformer head with a class token
-    to a 2-way classifier. FP32, deterministic first-64600-sample (~4.04 s @ 16 kHz) window, tile-repeat
-    if shorter (no random crop, no resampling). score = softmax(logits)[bonafide]; higher = more bona
-    fide. Official Speech-Arena-2025/DF_Arena_1B_V_1 checkpoint.
+  description: RAPTOR universal anti-spoofing model. A wav2vec 2.0 XLS-R 1B self-supervised
+    front-end whose per-layer hidden states are combined by learnable attention pooling
+    (a layer-wise sigmoid gate over an attention-pooled summary), then passed through
+    a 4-block Conformer head with a class token to a 2-way classifier. FP32, deterministic
+    first-64600-sample (~4.04 s @ 16 kHz) window, tile-repeat if shorter (no random
+    crop, no resampling). score = softmax(logits)[bonafide]; higher = more bona fide.
+    Official Speech-Arena-2025/DF_Arena_1B_V_1 checkpoint.
   code: https://huggingface.co/Speech-Arena-2025/DF_Arena_1B_V_1
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DF_Arena_1B_V_1
   params_millions: 1147.8453
   paper:
     arxiv_id: '2603.06164'
     url: https://arxiv.org/abs/2603.06164
-    bibtex: "@misc{kulkarni2026compactsslbackbonesmatter,\n  title={Do Compact SSL Backbones Matter for\
-      \ Audio Deepfake Detection? A Controlled Study with RAPTOR},\n  author={Ajinkya Kulkarni and Sandipana\
-      \ Dowerah and Atharva Kulkarni and Tanel Alumäe and Mathew Magimai Doss},\n  year={2026},\n  eprint={2603.06164},\n\
-      \  archivePrefix={arXiv},\n  primaryClass={cs.SD},\n  url={https://arxiv.org/abs/2603.06164}\n}\n"
+    bibtex: "@misc{kulkarni2026compactsslbackbonesmatter,\n  title={Do Compact SSL\
+      \ Backbones Matter for Audio Deepfake Detection? A Controlled Study with RAPTOR},\n\
+      \  author={Ajinkya Kulkarni and Sandipana Dowerah and Atharva Kulkarni and Tanel\
+      \ Alumäe and Mathew Magimai Doss},\n  year={2026},\n  eprint={2603.06164},\n\
+      \  archivePrefix={arXiv},\n  primaryClass={cs.SD},\n  url={https://arxiv.org/abs/2603.06164}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -34,15 +37,17 @@ submitter:
   hf_username: korallll
   contact: kborodin.research@gmail.com
 submitted_at: '2026-06-25'
-notes: DF Arena 1B (RAPTOR family), XLS-R 1B (wav2vec 2.0) front-end + attention layer pooling + 4-block
-  Conformer back-end. Architecture is built from the facebook/wav2vec2-xls-r-1b config, then every weight
-  is overwritten by the fine-tuned checkpoint (pytorch_model.bin). Trained on traditional + singing +
-  environmental deepfake corpora (ASVspoof 2019/2024, Codecfake, LibriSeVoc, DFADD, CTRSVDD, SpoofCeleb,
-  MLAAD, EnvSDD). The vendored forward feeds a real (B,T) batch to the SSL model (source hard-coded batch-1
-  via unsqueeze(0)); numerically identical per item in eval() (BatchNorm uses running stats). Deterministic
-  first-64600-sample window. score = output softmax prob for class 1 (bona fide); higher = more bona fide.
+notes: DF Arena 1B (RAPTOR family), XLS-R 1B (wav2vec 2.0) front-end + attention layer
+  pooling + 4-block Conformer back-end. Architecture is built from the facebook/wav2vec2-xls-r-1b
+  config, then every weight is overwritten by the fine-tuned checkpoint (pytorch_model.bin).
+  Trained on traditional + singing + environmental deepfake corpora (ASVspoof 2019/2024,
+  Codecfake, LibriSeVoc, DFADD, CTRSVDD, SpoofCeleb, MLAAD, EnvSDD). The vendored
+  forward feeds a real (B,T) batch to the SSL model (source hard-coded batch-1 via
+  unsqueeze(0)); numerically identical per item in eval() (BatchNorm uses running
+  stats). Deterministic first-64600-sample window. score = output softmax prob for
+  class 1 (bona fide); higher = more bona fide.
 reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-25'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference

submissions/df-arena-500m.yaml

@@ -2,22 +2,25 @@ schema_version: 4
 system:
   name: DF Arena 500M
   slug: df-arena-500m
-  description: RAPTOR universal anti-spoofing model. A wav2vec 2.0 XLS-R 300M self-supervised front-end
-    whose per-layer hidden states are combined by learnable attention pooling (a layer-wise sigmoid gate
-    over an attention-pooled summary), then passed through a 4-block Conformer head with a class token
-    to a 2-way classifier. FP32, deterministic first-64600-sample (~4.04 s @ 16 kHz) window, tile-repeat
-    if shorter (no random crop, no resampling). score = softmax(logits)[bonafide]; higher = more bona
-    fide. Official Speech-Arena-2025/DF_Arena_500M_V_1 checkpoint.
+  description: RAPTOR universal anti-spoofing model. A wav2vec 2.0 XLS-R 300M self-supervised
+    front-end whose per-layer hidden states are combined by learnable attention pooling
+    (a layer-wise sigmoid gate over an attention-pooled summary), then passed through
+    a 4-block Conformer head with a class token to a 2-way classifier. FP32, deterministic
+    first-64600-sample (~4.04 s @ 16 kHz) window, tile-repeat if shorter (no random
+    crop, no resampling). score = softmax(logits)[bonafide]; higher = more bona fide.
+    Official Speech-Arena-2025/DF_Arena_500M_V_1 checkpoint.
   code: https://huggingface.co/Speech-Arena-2025/DF_Arena_500M_V_1
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DF_Arena_500M_V_1
   params_millions: 436.1919
   paper:
     arxiv_id: '2603.06164'
     url: https://arxiv.org/abs/2603.06164
-    bibtex: "@misc{kulkarni2026compactsslbackbonesmatter,\n  title={Do Compact SSL Backbones Matter for\
-      \ Audio Deepfake Detection? A Controlled Study with RAPTOR},\n  author={Ajinkya Kulkarni and Sandipana\
-      \ Dowerah and Atharva Kulkarni and Tanel Alumäe and Mathew Magimai Doss},\n  year={2026},\n  eprint={2603.06164},\n\
-      \  archivePrefix={arXiv},\n  primaryClass={cs.SD},\n  url={https://arxiv.org/abs/2603.06164}\n}\n"
+    bibtex: "@misc{kulkarni2026compactsslbackbonesmatter,\n  title={Do Compact SSL\
+      \ Backbones Matter for Audio Deepfake Detection? A Controlled Study with RAPTOR},\n\
+      \  author={Ajinkya Kulkarni and Sandipana Dowerah and Atharva Kulkarni and Tanel\
+      \ Alumäe and Mathew Magimai Doss},\n  year={2026},\n  eprint={2603.06164},\n\
+      \  archivePrefix={arXiv},\n  primaryClass={cs.SD},\n  url={https://arxiv.org/abs/2603.06164}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -34,15 +37,17 @@ submitter:
   hf_username: korallll
   contact: kborodin.research@gmail.com
 submitted_at: '2026-06-25'
-notes: DF Arena 500M (RAPTOR family), XLS-R 300M (wav2vec 2.0) front-end + attention layer pooling + 4-block
-  Conformer back-end. Architecture is built from the facebook/wav2vec2-xls-r-300m config, then every weight
-  is overwritten by the fine-tuned checkpoint (pytorch_model.bin). Trained on traditional + singing +
-  environmental deepfake corpora (ASVspoof 2019/2024, Codecfake, LibriSeVoc, DFADD, CTRSVDD, SpoofCeleb,
-  MLAAD, EnvSDD). The vendored forward feeds a real (B,T) batch to the SSL model (source hard-coded batch-1
-  via unsqueeze(0)); numerically identical per item in eval() (BatchNorm uses running stats). Deterministic
-  first-64600-sample window. score = output softmax prob for class 1 (bona fide); higher = more bona fide.
+notes: DF Arena 500M (RAPTOR family), XLS-R 300M (wav2vec 2.0) front-end + attention
+  layer pooling + 4-block Conformer back-end. Architecture is built from the facebook/wav2vec2-xls-r-300m
+  config, then every weight is overwritten by the fine-tuned checkpoint (pytorch_model.bin).
+  Trained on traditional + singing + environmental deepfake corpora (ASVspoof 2019/2024,
+  Codecfake, LibriSeVoc, DFADD, CTRSVDD, SpoofCeleb, MLAAD, EnvSDD). The vendored
+  forward feeds a real (B,T) batch to the SSL model (source hard-coded batch-1 via
+  unsqueeze(0)); numerically identical per item in eval() (BatchNorm uses running
+  stats). Deterministic first-64600-sample window. score = output softmax prob for
+  class 1 (bona fide); higher = more bona fide.
 reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-25'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference

submissions/nes2net.yaml

@@ -2,22 +2,25 @@ schema_version: 4
 system:
   name: Nes2Net
   slug: nes2net
-  description: wav2vec 2.0 (XLS-R 300M) self-supervised front-end fine-tuned end-to-end with a Nes2Net-X
-    (Nested Res2Net TDNN) back-end for speech anti-spoofing. The nested Res2Net structure couples multi-scale
-    residual groups with squeeze- excitation, replacing dimensionality-reducing necks; mean temporal pooling
-    + linear classifier. Only ~0.51M back-end params. Official Nes2Net-X single checkpoint (ASVspoof2021
-    LA 1.73% / DF 1.65% EER as reported), trained on ASVspoof2019 LA with RawBoost, FP32, deterministic
-    first-64600-sample window (no random crop).
+  description: wav2vec 2.0 (XLS-R 300M) self-supervised front-end fine-tuned end-to-end
+    with a Nes2Net-X (Nested Res2Net TDNN) back-end for speech anti-spoofing. The
+    nested Res2Net structure couples multi-scale residual groups with squeeze- excitation,
+    replacing dimensionality-reducing necks; mean temporal pooling + linear classifier.
+    Only ~0.51M back-end params. Official Nes2Net-X single checkpoint (ASVspoof2021
+    LA 1.73% / DF 1.65% EER as reported), trained on ASVspoof2019 LA with RawBoost,
+    FP32, deterministic first-64600-sample window (no random crop).
   code: https://github.com/Liu-Tianchi/Nes2Net_ASVspoof_ITW
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Nes2Net
   params_millions: 317.9026
   paper:
     arxiv_id: '2504.05657'
     url: https://arxiv.org/abs/2504.05657
-    bibtex: "@article{Nes2Net,\n  author={Liu, Tianchi and Truong, Duc-Tuan and Das, Rohan Kumar and Lee,\
-      \ Kong Aik and Li, Haizhou},\n  journal={IEEE Transactions on Information Forensics and Security},\n\
-      \  title={Nes2Net: A Lightweight Nested Architecture for Foundation Model Driven Speech Anti-Spoofing},\n\
-      \  year={2025},\n  volume={20},\n  pages={12005--12018},\n  doi={10.1109/TIFS.2025.3626963}\n}\n"
+    bibtex: "@article{Nes2Net,\n  author={Liu, Tianchi and Truong, Duc-Tuan and Das,\
+      \ Rohan Kumar and Lee, Kong Aik and Li, Haizhou},\n  journal={IEEE Transactions\
+      \ on Information Forensics and Security},\n  title={Nes2Net: A Lightweight Nested\
+      \ Architecture for Foundation Model Driven Speech Anti-Spoofing},\n  year={2025},\n\
+      \  volume={20},\n  pages={12005--12018},\n  doi={10.1109/TIFS.2025.3626963}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -34,15 +37,16 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: XLS-R 300M (wav2vec 2.0) front-end + Nes2Net-X back-end, the single (non-averaged) checkpoint from
-  Liu-Tianchi/Nes2Net_ASVspoof_ITW (Nes_ratio [8,8], SE_ratio [1], pool_func 'mean', dilation 2). Architecture
-  is built from the base xlsr2_300m.pt model config, then every weight is overwritten by the fine-tuned
-  checkpoint. Deterministic first-64600-sample window (no random crop), matching the source data_utils_SSL.py::pad
-  used at eval (default --test_protocol 4sec). score = output logit for class 1 (bona fide); higher =
-  more bona fide. Back-end params ~0.51M; params_millions reports the full deployed model incl. the XLS-R
-  front-end.
+notes: XLS-R 300M (wav2vec 2.0) front-end + Nes2Net-X back-end, the single (non-averaged)
+  checkpoint from Liu-Tianchi/Nes2Net_ASVspoof_ITW (Nes_ratio [8,8], SE_ratio [1],
+  pool_func 'mean', dilation 2). Architecture is built from the base xlsr2_300m.pt
+  model config, then every weight is overwritten by the fine-tuned checkpoint. Deterministic
+  first-64600-sample window (no random crop), matching the source data_utils_SSL.py::pad
+  used at eval (default --test_protocol 4sec). score = output logit for class 1 (bona
+  fide); higher = more bona fide. Back-end params ~0.51M; params_millions reports
+  the full deployed model incl. the XLS-R front-end.

submissions/random-baseline.yaml

@@ -1,65 +1,52 @@
 schema_version: 4
-
 system:
   name: random-baseline
   slug: random-baseline
-  description: >
-    Reference random baseline. Returns N(0, 1) for every utterance using a
-    fixed seed (seed=0). EER ≈ 50% by construction. Used as the seeded
-    smoke-test baseline for the arena (roadmap Phase 3).
+  description: 'Reference random baseline. Returns N(0, 1) for every utterance using
+    a fixed seed (seed=0). EER ≈ 50% by construction. Used as the seeded smoke-test
+    baseline for the arena (roadmap Phase 3).
+
+    '
   code: https://github.com/SpeechAntiSpoofingBenchmarks/speech-spoof-bench
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/random-baseline-asas
   params_millions: 1
   paper:
-    arxiv_id: "1911.01601"
+    arxiv_id: '1911.01601'
     url: https://arxiv.org/abs/1911.01601
-    bibtex: |
-      @article{wang2020asvspoof,
-        title={ASVspoof 2019: A large-scale public database of synthesized,
-               converted and replayed speech},
-        author={Wang, Xin and Yamagishi, Junichi and Todisco, Massimiliano and
-                Delgado, H{\'e}ctor and Nautsch, Andreas and Evans, Nicholas
-                and Sahidullah, Md and Vestman, Ville and Kinnunen, Tomi and
-                Lee, Kong Aik and others},
-        journal={Computer Speech \& Language},
-        volume={64},
-        pages={101114},
-        year={2020},
-        publisher={Elsevier}
-      }
-
+    bibtex: "@article{wang2020asvspoof,\n  title={ASVspoof 2019: A large-scale public\
+      \ database of synthesized,\n         converted and replayed speech},\n  author={Wang,\
+      \ Xin and Yamagishi, Junichi and Todisco, Massimiliano and\n          Delgado,\
+      \ H{\\'e}ctor and Nautsch, Andreas and Evans, Nicholas\n          and Sahidullah,\
+      \ Md and Vestman, Ville and Kinnunen, Tomi and\n          Lee, Kong Aik and\
+      \ others},\n  journal={Computer Speech \\& Language},\n  volume={64},\n  pages={101114},\n\
+      \  year={2020},\n  publisher={Elsevier}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
   split: test
-
 scores:
   eer_percent: 48.21192052980132
   n_trials: 3755
   n_skipped: 0
-
 artifact:
   scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/random-baseline-asas/resolve/78d28e1b7be5df39771b8f0f0151c68def153d58/.eval_results/SpeechAntiSpoofingBenchmarks/DFADD/scores.txt
   scores_sha256: e217aa5e91d2410fdb9baeb5de7a291aeaa3d844efaa472fece67edc8a214bc8
   bench_version: speech-spoof-bench==0.3.4
-
 reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: 2026-06-10
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
-
+  match: inference
 submitter:
   hf_username: SpeechAntiSpoofingBenchmarks
   contact: k.n.borodin@mtuci.ru
-
 submitted_at: 2026-06-10
-
-notes: >
-  Seeded random baseline for the newly-added DFADD dataset (DFADD: The Diffusion
-  and Flow-Matching Based Audio Deepfake Dataset, arXiv 2409.08731; MIT;
-  VCTK-derived; eval/test split). 3755 trials: 755 bonafide (VCTK) / 3000 spoof
-  from 5 modern TTS (Grad-TTS, Matcha-TTS, NaturalSpeech 2, PFlow-TTS, StyleTTS 2,
-  600 each). Random baseline has no paper of its own; the paper field cites the
-  ASVspoof 2019 dataset paper as a placeholder, per the Phase 3 convention and
-  consistent with the other random-baseline submissions (same system slug).
+notes: 'Seeded random baseline for the newly-added DFADD dataset (DFADD: The Diffusion
+  and Flow-Matching Based Audio Deepfake Dataset, arXiv 2409.08731; MIT; VCTK-derived;
+  eval/test split). 3755 trials: 755 bonafide (VCTK) / 3000 spoof from 5 modern TTS
+  (Grad-TTS, Matcha-TTS, NaturalSpeech 2, PFlow-TTS, StyleTTS 2, 600 each). Random
+  baseline has no paper of its own; the paper field cites the ASVspoof 2019 dataset
+  paper as a placeholder, per the Phase 3 convention and consistent with the other
+  random-baseline submissions (same system slug).
+
+  '

submissions/rawbmamba.yaml

@@ -2,9 +2,10 @@ schema_version: 4
 system:
   name: RawBMamba
   slug: rawbmamba
-  description: 'RawBMamba: end-to-end bidirectional state-space (Mamba) model for audio deepfake detection. Sinc/Res front-end
-    + bidirectional Mamba blocks with attention pooling and an angular-margin (A-softmax) head. Raw-waveform, FP32, deterministic
-    first-64600-sample (4 s) window. Higher score = more bona fide.
+  description: 'RawBMamba: end-to-end bidirectional state-space (Mamba) model for
+    audio deepfake detection. Sinc/Res front-end + bidirectional Mamba blocks with
+    attention pooling and an angular-margin (A-softmax) head. Raw-waveform, FP32,
+    deterministic first-64600-sample (4 s) window. Higher score = more bona fide.
 
     '
   code: https://github.com/cyjie429/RawBMamba
@@ -13,10 +14,12 @@ system:
   paper:
     arxiv_id: '2406.06086'
     url: https://arxiv.org/abs/2406.06086
-    bibtex: "@inproceedings{chen2024rawbmamba,\n  title     = {RawBMamba: End-to-End Bidirectional State Space Model for Audio\
-      \ Deepfake Detection},\n  author    = {Chen, Yujie and Yi, Jiangyan and Xue, Jianhua and Wang, Chenglong and Zhang,\
-      \ Xiaohui and Dong, Shunbo and Zeng, Siding and Tao, Jianhua and Zhao, Lv and Wang, Cunhang},\n  booktitle = {Interspeech\
-      \ 2024},\n  year      = {2024}\n}\n"
+    bibtex: "@inproceedings{chen2024rawbmamba,\n  title     = {RawBMamba: End-to-End\
+      \ Bidirectional State Space Model for Audio Deepfake Detection},\n  author \
+      \   = {Chen, Yujie and Yi, Jiangyan and Xue, Jianhua and Wang, Chenglong and\
+      \ Zhang, Xiaohui and Dong, Shunbo and Zeng, Siding and Tao, Jianhua and Zhao,\
+      \ Lv and Wang, Cunhang},\n  booktitle = {Interspeech 2024},\n  year      = {2024}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -33,10 +36,10 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-26'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-26'
-notes: Pretrained rawbmamba_best.pt. Mamba selective-scan run on the pure-PyTorch reference path (CUDA kernels optional).
-  Deterministic first-64600-sample window.
+notes: Pretrained rawbmamba_best.pt. Mamba selective-scan run on the pure-PyTorch
+  reference path (CUDA kernels optional). Deterministic first-64600-sample window.

submissions/rawtfnet.yaml

@@ -2,18 +2,19 @@ schema_version: 4
 system:
   name: RawTFNet
   slug: rawtfnet
-  description: 'Lightweight raw-waveform CNN for speech anti-spoofing: sinc-convolution front-end, depthwise-separable
-    Res2Net-SE conv blocks, and a Tf-SepNet (time-frequency separable) classifier (depth=10, width=32).
-    ASVspoof2019 LA pretrained, FP32, deterministic first-64600-sample window (no random crop).'
+  description: 'Lightweight raw-waveform CNN for speech anti-spoofing: sinc-convolution
+    front-end, depthwise-separable Res2Net-SE conv blocks, and a Tf-SepNet (time-frequency
+    separable) classifier (depth=10, width=32). ASVspoof2019 LA pretrained, FP32,
+    deterministic first-64600-sample window (no random crop).'
   code: https://github.com/swagshaw/RawTFNet-Pytorch
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/RawTFNet/blob/aa12f0fe2f10cc5278c954175a12c18cc43e3113/Best_RawTFNet_32.pth
   params_millions: 0.17754
   paper:
     arxiv_id: '2507.08227'
     url: https://arxiv.org/abs/2507.08227
-    bibtex: "@article{xiao2025rawtfnet,\n  title={RawTFNet: A Lightweight CNN Architecture for Speech\
-      \ Anti-spoofing},\n  author={Xiao, Yang and Dang, Ting and Das, Rohan Kumar},\n  journal={arXiv\
-      \ preprint arXiv:2507.08227},\n  year={2025}\n}\n"
+    bibtex: "@article{xiao2025rawtfnet,\n  title={RawTFNet: A Lightweight CNN Architecture\
+      \ for Speech Anti-spoofing},\n  author={Xiao, Yang and Dang, Ting and Das, Rohan\
+      \ Kumar},\n  journal={arXiv preprint arXiv:2507.08227},\n  year={2025}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -30,9 +31,10 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: Deterministic first-64600-sample window (no random crop). Full RawTFNet (width=32), Best_RawTFNet_32.pth.
+notes: Deterministic first-64600-sample window (no random crop). Full RawTFNet (width=32),
+  Best_RawTFNet_32.pth.

submissions/res2tcnguard.yaml

@@ -2,21 +2,22 @@ schema_version: 4
 system:
   name: Res2TCNGuard
   slug: res2tcnguard
-  description: 'TCN-based anti-spoofing countermeasure: sinc-convolution front-end, Res2Net encoder, and
-    dual temporal convolutional networks. ASVspoof2019 LA pretrained, FP32, deterministic first-64600-sample
-    window (no random crop).'
+  description: 'TCN-based anti-spoofing countermeasure: sinc-convolution front-end,
+    Res2Net encoder, and dual temporal convolutional networks. ASVspoof2019 LA pretrained,
+    FP32, deterministic first-64600-sample window (no random crop).'
   code: https://github.com/lab260ru/Res2TCNGuard
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Res2TCNGuard/blob/523ef915d798df3deee2be4fd06af7e1965ede14/best_1.495.pth
   params_millions: 0.172102
   paper:
     arxiv_id: 10.48084/etasr.8906
     url: https://etasr.com/index.php/ETASR/article/view/8906
-    bibtex: "@article{Borodin_Kudryavtsev_Mkrtchian_Gorodnichev_2024,\n  place={Greece},\n  title={Capsule-based\
-      \ and TCN-based Approaches for Spoofing Detection in Voice Biometry},\n  volume={14},\n  number={6},\n\
-      \  url={https://etasr.com/index.php/ETASR/article/view/8906},\n  DOI={10.48084/etasr.8906},\n  journal={Engineering,\
-      \ Technology & Applied Science Research},\n  author={Borodin, Kirill and Kudryavtsev, Vasiliy and\
-      \ Mkrtchian, Grach and Gorodnichev, Mikhail},\n  year={2024},\n  month={Dec.},\n  pages={18409--18414}\n\
-      }\n"
+    bibtex: "@article{Borodin_Kudryavtsev_Mkrtchian_Gorodnichev_2024,\n  place={Greece},\n\
+      \  title={Capsule-based and TCN-based Approaches for Spoofing Detection in Voice\
+      \ Biometry},\n  volume={14},\n  number={6},\n  url={https://etasr.com/index.php/ETASR/article/view/8906},\n\
+      \  DOI={10.48084/etasr.8906},\n  journal={Engineering, Technology & Applied\
+      \ Science Research},\n  author={Borodin, Kirill and Kudryavtsev, Vasiliy and\
+      \ Mkrtchian, Grach and Gorodnichev, Mikhail},\n  year={2024},\n  month={Dec.},\n\
+      \  pages={18409--18414}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -33,7 +34,7 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru

submissions/rescapsguard.yaml

@@ -2,21 +2,23 @@ schema_version: 4
 system:
   name: ResCapsGuard
   slug: rescapsguard
-  description: 'Capsule-based anti-spoofing countermeasure: sinc-convolution front-end, Res2Net-style
-    encoder, primary capsules, and dynamic routing to two output capsules (bona fide vs. spoof). ASVspoof2019
-    LA pretrained, FP32, deterministic first-64600-sample window (no random crop).'
+  description: 'Capsule-based anti-spoofing countermeasure: sinc-convolution front-end,
+    Res2Net-style encoder, primary capsules, and dynamic routing to two output capsules
+    (bona fide vs. spoof). ASVspoof2019 LA pretrained, FP32, deterministic first-64600-sample
+    window (no random crop).'
   code: https://github.com/lab260ru/ResCapsGuard
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/ResCapsGuard/blob/4f7f2bc745cb3980f9fd3c29811afe0744a28edd/new_capsules_changed_sinc_layer.pth
   params_millions: 1.606664
   paper:
     arxiv_id: 10.48084/etasr.8906
     url: https://etasr.com/index.php/ETASR/article/view/8906
-    bibtex: "@article{Borodin_Kudryavtsev_Mkrtchian_Gorodnichev_2024,\n  place={Greece},\n  title={Capsule-based\
-      \ and TCN-based Approaches for Spoofing Detection in Voice Biometry},\n  volume={14},\n  number={6},\n\
-      \  url={https://etasr.com/index.php/ETASR/article/view/8906},\n  DOI={10.48084/etasr.8906},\n  journal={Engineering,\
-      \ Technology & Applied Science Research},\n  author={Borodin, Kirill and Kudryavtsev, Vasiliy and\
-      \ Mkrtchian, Grach and Gorodnichev, Mikhail},\n  year={2024},\n  month={Dec.},\n  pages={18409--18414}\n\
-      }\n"
+    bibtex: "@article{Borodin_Kudryavtsev_Mkrtchian_Gorodnichev_2024,\n  place={Greece},\n\
+      \  title={Capsule-based and TCN-based Approaches for Spoofing Detection in Voice\
+      \ Biometry},\n  volume={14},\n  number={6},\n  url={https://etasr.com/index.php/ETASR/article/view/8906},\n\
+      \  DOI={10.48084/etasr.8906},\n  journal={Engineering, Technology & Applied\
+      \ Science Research},\n  author={Borodin, Kirill and Kudryavtsev, Vasiliy and\
+      \ Mkrtchian, Grach and Gorodnichev, Mikhail},\n  year={2024},\n  month={Dec.},\n\
+      \  pages={18409--18414}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -33,7 +35,7 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru

submissions/sonar-full.yaml

@@ -2,24 +2,26 @@ schema_version: 4
 system:
   name: SONAR-Full
   slug: sonar-full
-  description: 'SONAR (Spectral-Contrastive Audio Residuals): a dual-path detector with two wav2vec 2.0
-    XLS-R 300M encoders. One encodes the raw waveform (content / low-frequency path); the other encodes
-    a high-frequency residual from a learnable constrained SRM filter bank (30 filters, k=5). The two
-    feature streams are fused by bidirectional cross-attention, projected to 128-d, and classified by
-    an AASIST spectro-temporal graph-attention back-end; training adds a Jensen-Shannon alignment loss
-    between the two paths. Official SONAR-Full checkpoint (sonar_full_xlsr_aasist_eer6.pth), trained on
-    ASVspoof2019 LA, FP32, deterministic first-64600-sample window (zero-padded).'
+  description: 'SONAR (Spectral-Contrastive Audio Residuals): a dual-path detector
+    with two wav2vec 2.0 XLS-R 300M encoders. One encodes the raw waveform (content
+    / low-frequency path); the other encodes a high-frequency residual from a learnable
+    constrained SRM filter bank (30 filters, k=5). The two feature streams are fused
+    by bidirectional cross-attention, projected to 128-d, and classified by an AASIST
+    spectro-temporal graph-attention back-end; training adds a Jensen-Shannon alignment
+    loss between the two paths. Official SONAR-Full checkpoint (sonar_full_xlsr_aasist_eer6.pth),
+    trained on ASVspoof2019 LA, FP32, deterministic first-64600-sample window (zero-padded).'
   code: https://github.com/idonithid/SONAR-Audio-DF-Detection
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SONAR-Full
   params_millions: 641.5252
   paper:
     arxiv_id: '2511.21325'
     url: https://proceedings.mlr.press/v306/hidekel26a.html
-    bibtex: "@inproceedings{hidekel2026sonar,\n  title     = {{SONAR}: Spectral-Contrastive Audio Residuals\
-      \ for Generalizable Deepfake Detection},\n  author    = {Hidekel, Ido Nitzan and Lifshitz, Gal and\
-      \ Cohen, Khen and Raviv, Dan},\n  booktitle = {Proceedings of the 43rd International Conference\
-      \ on Machine Learning (ICML)},\n  series    = {Proceedings of Machine Learning Research},\n  volume\
-      \    = {306},\n  pages     = {42998--43012},\n  year      = {2026}\n}\n"
+    bibtex: "@inproceedings{hidekel2026sonar,\n  title     = {{SONAR}: Spectral-Contrastive\
+      \ Audio Residuals for Generalizable Deepfake Detection},\n  author    = {Hidekel,\
+      \ Ido Nitzan and Lifshitz, Gal and Cohen, Khen and Raviv, Dan},\n  booktitle\
+      \ = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},\n\
+      \  series    = {Proceedings of Machine Learning Research},\n  volume    = {306},\n\
+      \  pages     = {42998--43012},\n  year      = {2026}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -36,13 +38,14 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-10-28'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-10-28'
-notes: 'SONAR-Full from idonithid/SONAR-Audio-DF-Detection (ICML 2026): dual XLS-R 300M encoders (raw
-  waveform + constrained-SRM high-frequency residual) fused by cross-attention, AASIST back-end. Official
-  sonar_full_xlsr_aasist_eer6.pth weights (CC-BY-NC-4.0), strict load. Deterministic first-64600-sample
-  window, zero-padded for shorter clips (source sonar/inference.py). score = output logit for class 1
+notes: 'SONAR-Full from idonithid/SONAR-Audio-DF-Detection (ICML 2026): dual XLS-R
+  300M encoders (raw waveform + constrained-SRM high-frequency residual) fused by
+  cross-attention, AASIST back-end. Official sonar_full_xlsr_aasist_eer6.pth weights
+  (CC-BY-NC-4.0), strict load. Deterministic first-64600-sample window, zero-padded
+  for shorter clips (source sonar/inference.py). score = output logit for class 1
   (real); higher = more bona fide.'

submissions/spectra-0.yaml

@@ -2,10 +2,11 @@ schema_version: 4
 system:
   name: Spectra-0
   slug: spectra-0
-  description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised front-end feeding a
-    single-layer MLP bridge (1024->128, SELU) into an ECAPA-TDNN 2-class classifier. Pre-release lab260/spectra_0
-    checkpoint. FP32, preemphasis (0.97), deterministic first-64600-sample window (no random crop). score
-    = output logit for class 1 (bona fide). Unpublished / pre-release model (no paper).
+  description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised
+    front-end feeding a single-layer MLP bridge (1024->128, SELU) into an ECAPA-TDNN
+    2-class classifier. Pre-release lab260/spectra_0 checkpoint. FP32, preemphasis
+    (0.97), deterministic first-64600-sample window (no random crop). score = output
+    logit for class 1 (bona fide). Unpublished / pre-release model (no paper).
 
     '
   code: https://huggingface.co/lab260/spectra_0
@@ -27,11 +28,11 @@ submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-24'
-notes: Spectra-0 (wav2vec2 XLS-R-300m SSL front-end + MLP bridge + ECAPA-TDNN). Preemphasis 0.97, deterministic
-  first-64600-sample window. score = logit class 1 (bona fide). Unpublished/pre-release (no paper). Scores
-  hosted in lab260/spectra_0/.eval_results.
+notes: Spectra-0 (wav2vec2 XLS-R-300m SSL front-end + MLP bridge + ECAPA-TDNN). Preemphasis
+  0.97, deterministic first-64600-sample window. score = logit class 1 (bona fide).
+  Unpublished/pre-release (no paper). Scores hosted in lab260/spectra_0/.eval_results.
 reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-24'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference

submissions/spectra-aasist.yaml

@@ -2,11 +2,12 @@ schema_version: 4
 system:
   name: Spectra-AASIST
   slug: spectra-aasist
-  description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised front-end feeding a
-    single-layer MLP bridge (1024->128) into an AASIST 2-class classifier with four-branch spectro-temporal
-    graph attention. Pre-release lab260/Spectra-AASIST checkpoint. FP32, preemphasis (0.97), deterministic
-    first-64600-sample window (no random crop). score = output logit for class 1 (bona fide). Unpublished
-    / pre-release model (no paper).
+  description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised
+    front-end feeding a single-layer MLP bridge (1024->128) into an AASIST 2-class
+    classifier with four-branch spectro-temporal graph attention. Pre-release lab260/Spectra-AASIST
+    checkpoint. FP32, preemphasis (0.97), deterministic first-64600-sample window
+    (no random crop). score = output logit for class 1 (bona fide). Unpublished /
+    pre-release model (no paper).
 
     '
   code: https://huggingface.co/lab260/Spectra-AASIST
@@ -28,11 +29,11 @@ submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-24'
-notes: Spectra-AASIST (wav2vec2 XLS-R-300m SSL front-end + MLP bridge + AASIST). Preemphasis 0.97, deterministic
-  first-64600-sample window. score = logit class 1 (bona fide). Unpublished/pre-release (no paper). Scores
-  hosted in lab260/Spectra-AASIST/.eval_results.
+notes: Spectra-AASIST (wav2vec2 XLS-R-300m SSL front-end + MLP bridge + AASIST). Preemphasis
+  0.97, deterministic first-64600-sample window. score = logit class 1 (bona fide).
+  Unpublished/pre-release (no paper). Scores hosted in lab260/Spectra-AASIST/.eval_results.
 reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-24'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference

submissions/spectra-aasist3.yaml

@@ -2,11 +2,12 @@ schema_version: 4
 system:
   name: Spectra-AASIST3
   slug: spectra-aasist3
-  description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised front-end feeding a
-    single-layer MLP bridge (1024->128) into a KAN-enhanced AASIST (KAN-AASIST) 2-class classifier with
-    four-branch spectro-temporal graph attention and a KAN output layer. Pre-release lab260/Spectra-AASIST3
-    checkpoint. FP32, preemphasis (0.97), deterministic first-64600-sample window (no random crop). score
-    = output logit for class 1 (bona fide). Unpublished / pre-release model (no paper).
+  description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised
+    front-end feeding a single-layer MLP bridge (1024->128) into a KAN-enhanced AASIST
+    (KAN-AASIST) 2-class classifier with four-branch spectro-temporal graph attention
+    and a KAN output layer. Pre-release lab260/Spectra-AASIST3 checkpoint. FP32, preemphasis
+    (0.97), deterministic first-64600-sample window (no random crop). score = output
+    logit for class 1 (bona fide). Unpublished / pre-release model (no paper).
 
     '
   code: https://huggingface.co/lab260/Spectra-AASIST3
@@ -28,11 +29,11 @@ submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-24'
-notes: Spectra-AASIST3 (wav2vec2 XLS-R-300m SSL front-end + MLP bridge + KAN-AASIST). Preemphasis 0.97,
-  deterministic first-64600-sample window. score = logit class 1 (bona fide). Unpublished/pre-release
-  (no paper). Scores hosted in lab260/Spectra-AASIST3/.eval_results.
+notes: Spectra-AASIST3 (wav2vec2 XLS-R-300m SSL front-end + MLP bridge + KAN-AASIST).
+  Preemphasis 0.97, deterministic first-64600-sample window. score = logit class 1
+  (bona fide). Unpublished/pre-release (no paper). Scores hosted in lab260/Spectra-AASIST3/.eval_results.
 reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-24'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference

submissions/ssl-feature-gating.yaml

@@ -2,10 +2,11 @@ schema_version: 4
 system:
   name: SSL Feature Gating
   slug: ssl-feature-gating
-  description: 'Multi-level SSL Feature Gating for audio deepfake detection: a gating mechanism over XLS-R (wav2vec2-xls-r-300m)
-    hidden states as front-end, Multi-kernel gated Convolution (MultiConv) back-end with CKA-enforced feature diversity, and
-    multi-head attention pooling. Raw-waveform, FP32, deterministic first-64600-sample (4 s) window. Higher score = more bona
-    fide.
+  description: 'Multi-level SSL Feature Gating for audio deepfake detection: a gating
+    mechanism over XLS-R (wav2vec2-xls-r-300m) hidden states as front-end, Multi-kernel
+    gated Convolution (MultiConv) back-end with CKA-enforced feature diversity, and
+    multi-head attention pooling. Raw-waveform, FP32, deterministic first-64600-sample
+    (4 s) window. Higher score = more bona fide.
 
     '
   code: https://github.com/hoanmyTran/dissimilarity_deepfake_detection
@@ -14,10 +15,12 @@ system:
   paper:
     arxiv_id: 10.1145/3746027.3754568
     url: https://dl.acm.org/doi/10.1145/3746027.3754568
-    bibtex: "@inproceedings{tran2025ssl,\n  title     = {Multi-level SSL Feature Gating for Audio Deepfake Detection},\n \
-      \ author    = {Tran, Hoan My and Lolive, Damien and Sini, Aghilas and Delhay, Arnaud and Marteau, Pierre-Fran\\c{c}ois\
-      \ and Guennec, David},\n  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia (MM '25)},\n\
-      \  year      = {2025},\n  doi       = {10.1145/3746027.3754568}\n}\n"
+    bibtex: "@inproceedings{tran2025ssl,\n  title     = {Multi-level SSL Feature Gating\
+      \ for Audio Deepfake Detection},\n  author    = {Tran, Hoan My and Lolive, Damien\
+      \ and Sini, Aghilas and Delhay, Arnaud and Marteau, Pierre-Fran\\c{c}ois and\
+      \ Guennec, David},\n  booktitle = {Proceedings of the 33rd ACM International\
+      \ Conference on Multimedia (MM '25)},\n  year      = {2025},\n  doi       =\
+      \ {10.1145/3746027.3754568}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -34,10 +37,10 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-26'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-26'
-notes: ACM MM 2025 checkpoint acmmm25_best_model.ckpt. XLS-R-300M front-end + MultiConv gating. Deterministic first-64600-sample
-  window.
+notes: ACM MM 2025 checkpoint acmmm25_best_model.ckpt. XLS-R-300M front-end + MultiConv
+  gating. Deterministic first-64600-sample window.

submissions/w2v2-aasist.yaml

@@ -2,22 +2,23 @@ schema_version: 4
 system:
   name: W2V2-AASIST
   slug: w2v2-aasist
-  description: wav2vec 2.0 (XLS-R 300M) self-supervised front-end fine-tuned end-to-end with an AASIST
-    spectro-temporal graph-attention back-end for speech anti-spoofing. The XLS-R features are projected
-    to 128-d, max-pooled, and fed through a RawNet2-style residual encoder and heterogeneous stacking
-    graph-attention layers. Official TakHemlata/SSL_Anti-spoofing LA checkpoint (LA_model.pth), trained
-    on ASVspoof2019 LA with RawBoost augmentation, FP32, deterministic first-64600-sample window (no random
-    crop).
+  description: wav2vec 2.0 (XLS-R 300M) self-supervised front-end fine-tuned end-to-end
+    with an AASIST spectro-temporal graph-attention back-end for speech anti-spoofing.
+    The XLS-R features are projected to 128-d, max-pooled, and fed through a RawNet2-style
+    residual encoder and heterogeneous stacking graph-attention layers. Official TakHemlata/SSL_Anti-spoofing
+    LA checkpoint (LA_model.pth), trained on ASVspoof2019 LA with RawBoost augmentation,
+    FP32, deterministic first-64600-sample window (no random crop).
   code: https://github.com/TakHemlata/SSL_Anti-spoofing
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/W2V2-AASIST
   params_millions: 317.8378
   paper:
     arxiv_id: '2202.12233'
     url: https://arxiv.org/abs/2202.12233
-    bibtex: "@inproceedings{tak2022automatic,\n  title={Automatic speaker verification spoofing and deepfake\
-      \ detection using wav2vec 2.0 and data augmentation},\n  author={Tak, Hemlata and Todisco, Massimiliano\
-      \ and Wang, Xin and Jung, Jee-weon and Yamagishi, Junichi and Evans, Nicholas},\n  booktitle={The\
-      \ Speaker and Language Recognition Workshop (Odyssey 2022)},\n  pages={112--119},\n  year={2022}\n\
+    bibtex: "@inproceedings{tak2022automatic,\n  title={Automatic speaker verification\
+      \ spoofing and deepfake detection using wav2vec 2.0 and data augmentation},\n\
+      \  author={Tak, Hemlata and Todisco, Massimiliano and Wang, Xin and Jung, Jee-weon\
+      \ and Yamagishi, Junichi and Evans, Nicholas},\n  booktitle={The Speaker and\
+      \ Language Recognition Workshop (Odyssey 2022)},\n  pages={112--119},\n  year={2022}\n\
       }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
@@ -35,13 +36,14 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: XLS-R 300M (wav2vec 2.0) front-end + AASIST back-end ("W2V2-AASIST"), the LA variant (LA_model.pth)
-  from TakHemlata/SSL_Anti-spoofing. Architecture is built from the base xlsr2_300m.pt model config, then
-  every weight is overwritten by the fine-tuned checkpoint. Deterministic first-64600-sample window (no
-  random crop), matching the source data_utils_SSL.py::pad used at eval. score = output logit for class
-  1 (bona fide); higher = more bona fide.
+notes: XLS-R 300M (wav2vec 2.0) front-end + AASIST back-end ("W2V2-AASIST"), the LA
+  variant (LA_model.pth) from TakHemlata/SSL_Anti-spoofing. Architecture is built
+  from the base xlsr2_300m.pt model config, then every weight is overwritten by the
+  fine-tuned checkpoint. Deterministic first-64600-sample window (no random crop),
+  matching the source data_utils_SSL.py::pad used at eval. score = output logit for
+  class 1 (bona fide); higher = more bona fide.

submissions/whisper-mfcc-mesonet.yaml

@@ -2,10 +2,11 @@ schema_version: 4
 system:
   name: Whisper-MFCC-MesoNet
   slug: whisper-mfcc-mesonet
-  description: '(Whisper + MFCC) MesoNet anti-spoofing countermeasure: a Whisper tiny.en audio encoder
-    front-end concatenated (2 channels) with an MFCC+Δ+ΔΔ front-end, feeding a MesoInception4 classifier.
-    Whisper encoder fine-tuned end-to-end. FP32. Upstream eval pipeline reproduced: sox silence-trim (silence
-    1 0.2 1% -1 0.2 1%) then a 30 s (480000-sample) repeat-pad window. This is the best MesoNet configuration
+  description: '(Whisper + MFCC) MesoNet anti-spoofing countermeasure: a Whisper tiny.en
+    audio encoder front-end concatenated (2 channels) with an MFCC+Δ+ΔΔ front-end,
+    feeding a MesoInception4 classifier. Whisper encoder fine-tuned end-to-end. FP32.
+    Upstream eval pipeline reproduced: sox silence-trim (silence 1 0.2 1% -1 0.2 1%)
+    then a 30 s (480000-sample) repeat-pad window. This is the best MesoNet configuration
     from the paper.'
   code: https://github.com/piotrkawa/deepfake-whisper-features
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WhisperMFCCMesoNet
@@ -13,10 +14,11 @@ system:
   paper:
     arxiv_id: '2306.01428'
     url: https://arxiv.org/abs/2306.01428
-    bibtex: "@inproceedings{kawa23b_interspeech,\n  title     = {Improved DeepFake Detection Using Whisper\
-      \ Features},\n  author    = {Piotr Kawa and Marcin Plata and Micha{\\l} Czuba and Piotr Szyma{\\\
-      'n}ski and Piotr Syga},\n  year      = {2023},\n  booktitle = {Proc. INTERSPEECH 2023},\n  pages\
-      \     = {4009--4013},\n  doi       = {10.21437/Interspeech.2023-1537},\n}\n"
+    bibtex: "@inproceedings{kawa23b_interspeech,\n  title     = {Improved DeepFake\
+      \ Detection Using Whisper Features},\n  author    = {Piotr Kawa and Marcin Plata\
+      \ and Micha{\\l} Czuba and Piotr Szyma{\\'n}ski and Piotr Syga},\n  year   \
+      \   = {2023},\n  booktitle = {Proc. INTERSPEECH 2023},\n  pages     = {4009--4013},\n\
+      \  doi       = {10.21437/Interspeech.2023-1537},\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -33,10 +35,11 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: Fine-tuned Whisper+MFCC MesoNet (best MesoNet config from the paper). Reproduces the paper's In-the-Wild
-  EER (26.72%) to within 0.01 pp with the upstream sox silence-trim + 30 s repeat-pad preprocessing.
+notes: Fine-tuned Whisper+MFCC MesoNet (best MesoNet config from the paper). Reproduces
+  the paper's In-the-Wild EER (26.72%) to within 0.01 pp with the upstream sox silence-trim
+  + 30 s repeat-pad preprocessing.

submissions/xlsr-mambo.yaml

@@ -2,11 +2,12 @@ schema_version: 4
 system:
   name: XLSR-MamBo
   slug: xlsr-mambo
-  description: 'XLSR-MamBo (MamBo-3, Hydra-N3): a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end (last
-    transformer layer) projected to 128-d, followed by a hybrid Mamba-attention backbone of 5 repeats
-    of [3x Hydra (bidirectional Mamba2/SSD quasiseparable mixer) + SwiGLU-MLP + multi-head attention +
-    SwiGLU-MLP] with pre-RMSNorm residuals, gated-attention pooling and a linear 2-way head. Official
-    MamBo-3 DA0 checkpoint (no RawBoost; trained on ASVspoof2019 LA), validation-loss best (best_0 / model_best),
+  description: 'XLSR-MamBo (MamBo-3, Hydra-N3): a fully fine-tuned wav2vec 2.0 XLS-R
+    300M front-end (last transformer layer) projected to 128-d, followed by a hybrid
+    Mamba-attention backbone of 5 repeats of [3x Hydra (bidirectional Mamba2/SSD quasiseparable
+    mixer) + SwiGLU-MLP + multi-head attention + SwiGLU-MLP] with pre-RMSNorm residuals,
+    gated-attention pooling and a linear 2-way head. Official MamBo-3 DA0 checkpoint
+    (no RawBoost; trained on ASVspoof2019 LA), validation-loss best (best_0 / model_best),
     FP32, deterministic first-66800-sample window (repeat-padded).'
   code: https://github.com/saki-ciallo/XLSR-MamBo
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/XLSR-MamBo
@@ -14,13 +15,14 @@ system:
   paper:
     arxiv_id: '2601.02944'
     url: https://aclanthology.org/2026.findings-acl.1573/
-    bibtex: "@inproceedings{ng-etal-2026-xlsr,\n  title     = {{XLSR}-{M}am{B}o: Scaling the Hybrid Mamba-Attention\
-      \ Backbone for Audio Deepfake Detection},\n  author    = {Ng, Kwok-Ho and Song, Tingting and Wu,\
-      \ Yongdong and Xia, Zhihua},\n  booktitle = {Findings of the Association for Computational Linguistics:\
-      \ ACL 2026},\n  month     = jul,\n  year      = {2026},\n  address   = {San Diego, California, United\
-      \ States},\n  publisher = {Association for Computational Linguistics},\n  pages     = {31450--31462},\n\
-      \  doi       = {10.18653/v1/2026.findings-acl.1573},\n  url       = {https://aclanthology.org/2026.findings-acl.1573/}\n\
-      }\n"
+    bibtex: "@inproceedings{ng-etal-2026-xlsr,\n  title     = {{XLSR}-{M}am{B}o: Scaling\
+      \ the Hybrid Mamba-Attention Backbone for Audio Deepfake Detection},\n  author\
+      \    = {Ng, Kwok-Ho and Song, Tingting and Wu, Yongdong and Xia, Zhihua},\n\
+      \  booktitle = {Findings of the Association for Computational Linguistics: ACL\
+      \ 2026},\n  month     = jul,\n  year      = {2026},\n  address   = {San Diego,\
+      \ California, United States},\n  publisher = {Association for Computational\
+      \ Linguistics},\n  pages     = {31450--31462},\n  doi       = {10.18653/v1/2026.findings-acl.1573},\n\
+      \  url       = {https://aclanthology.org/2026.findings-acl.1573/}\n}\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: c578c836da3b522b27d3dd85f89309f1737e5d31
@@ -37,15 +39,17 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-10-29'
   reproduced_bench_version: speech-spoof-bench==0.4.1
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-10-29'
-notes: 'XLSR-MamBo from saki-ciallo/XLSR-MamBo (ACL 2026 Findings), MamBo-3 (Hydra-N3) DA0 checkpoint
-  best_0.pt (= model_best.pt, lowest ASVspoof2019 LA validation loss), MIT, strict load. Deterministic
-  first-66800-sample window; shorter clips repeat-padded (source pad_wave_repeat). score = logit[bonafide]
-  - logit[spoof] (source eval_asv_txt.py). Deviations: waveform layer-norm is per utterance (torchaudio''s
-  bundle wrapper normalises jointly over the batch, making the source''s scores batch-dependent); FP32
-  instead of the source''s bf16 autocast; Mamba2 SSD scan and RMSNorms in plain PyTorch (identical math
-  to the mamba_ssm Triton kernels, max |score diff| 2e-4; Triton cannot compile them for sm_120).'
+notes: 'XLSR-MamBo from saki-ciallo/XLSR-MamBo (ACL 2026 Findings), MamBo-3 (Hydra-N3)
+  DA0 checkpoint best_0.pt (= model_best.pt, lowest ASVspoof2019 LA validation loss),
+  MIT, strict load. Deterministic first-66800-sample window; shorter clips repeat-padded
+  (source pad_wave_repeat). score = logit[bonafide] - logit[spoof] (source eval_asv_txt.py).
+  Deviations: waveform layer-norm is per utterance (torchaudio''s bundle wrapper normalises
+  jointly over the batch, making the source''s scores batch-dependent); FP32 instead
+  of the source''s bf16 autocast; Mamba2 SSD scan and RMSNorms in plain PyTorch (identical
+  math to the mamba_ssm Triton kernels, max |score diff| 2e-4; Triton cannot compile
+  them for sm_120).'

submissions/xlsr-sls.yaml

@@ -2,22 +2,25 @@ schema_version: 4
 system:
   name: XLSR-SLS
   slug: xlsr-sls
-  description: wav2vec 2.0 (XLS-R 300M) self-supervised front-end with the SLS (Sensitive Layer Selection)
-    classifier for audio deepfake detection. SLS gates and fuses the hidden states of all XLS-R transformer
-    layers — each layer contributing distinct discriminative cues — via a per-layer sigmoid attention,
-    sums the weighted multi-layer feature, then a BN + max-pool + two-layer MLP head emits a 2-way log-softmax.
-    Official QiShanZhang/SLSforASVspoof-2021-DF checkpoint (model_15, dev-EER 1.45%), trained on ASVspoof2019
-    LA, FP32, deterministic first-64600-sample window (no random crop).
+  description: wav2vec 2.0 (XLS-R 300M) self-supervised front-end with the SLS (Sensitive
+    Layer Selection) classifier for audio deepfake detection. SLS gates and fuses
+    the hidden states of all XLS-R transformer layers — each layer contributing distinct
+    discriminative cues — via a per-layer sigmoid attention, sums the weighted multi-layer
+    feature, then a BN + max-pool + two-layer MLP head emits a 2-way log-softmax.
+    Official QiShanZhang/SLSforASVspoof-2021-DF checkpoint (model_15, dev-EER 1.45%),
+    trained on ASVspoof2019 LA, FP32, deterministic first-64600-sample window (no
+    random crop).
   code: https://github.com/QiShanZhang/SLSforASVspoof-2021-DF
   checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/XLSR-SLS
   params_millions: 340.79
   paper:
     arxiv_id: 10.1145/3664647.3681345
     url: https://doi.org/10.1145/3664647.3681345
-    bibtex: "@inproceedings{zhang2024audio,\n  title={Audio Deepfake Detection with Self-Supervised XLS-R\
-      \ and SLS Classifier},\n  author={Zhang, Qishan and Wen, Shuangbing and Hu, Tao},\n  booktitle={Proceedings\
-      \ of the 32nd ACM International Conference on Multimedia},\n  pages={6765--6773},\n  year={2024},\n\
-      \  doi={10.1145/3664647.3681345}\n}\n"
+    bibtex: "@inproceedings{zhang2024audio,\n  title={Audio Deepfake Detection with\
+      \ Self-Supervised XLS-R and SLS Classifier},\n  author={Zhang, Qishan and Wen,\
+      \ Shuangbing and Hu, Tao},\n  booktitle={Proceedings of the 32nd ACM International\
+      \ Conference on Multimedia},\n  pages={6765--6773},\n  year={2024},\n  doi={10.1145/3664647.3681345}\n\
+      }\n"
 dataset:
   id: SpeechAntiSpoofingBenchmarks/DFADD
   revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
@@ -34,15 +37,17 @@ reproduction:
   reproduced_by: SpeechAntiSpoofingBenchmarks
   reproduced_at: '2026-06-10'
   reproduced_bench_version: speech-spoof-bench==0.3.4
-  match: scoring
+  match: inference
 submitter:
   hf_username: korallll
   contact: k.n.borodin@mtuci.ru
 submitted_at: '2026-06-10'
-notes: 'XLS-R 300M (wav2vec 2.0) front-end + SLS (Sensitive Layer Selection) classifier, from QiShanZhang/SLSforASVspoof-2021-DF
-  (ACM MM 2024). Architecture is built from the base xlsr2_300m.pt model config (shared with the W2V2-AASIST
-  submission), then every weight is overwritten by the fine-tuned checkpoint. SLS pools every transformer
-  layer''s hidden state, gates each by a learned sigmoid attention, and fuses them before a small MLP
-  head. Deterministic first-64600-sample window (no random crop); the head''s fc1 expects this fixed length.
-  score = log-softmax output for class 1 (bona fide); higher = more bona fide (source main.py: batch_score
-  = batch_out[:, 1]).'
+notes: 'XLS-R 300M (wav2vec 2.0) front-end + SLS (Sensitive Layer Selection) classifier,
+  from QiShanZhang/SLSforASVspoof-2021-DF (ACM MM 2024). Architecture is built from
+  the base xlsr2_300m.pt model config (shared with the W2V2-AASIST submission), then
+  every weight is overwritten by the fine-tuned checkpoint. SLS pools every transformer
+  layer''s hidden state, gates each by a learned sigmoid attention, and fuses them
+  before a small MLP head. Deterministic first-64600-sample window (no random crop);
+  the head''s fc1 expects this fixed length. score = log-softmax output for class
+  1 (bona fide); higher = more bona fide (source main.py: batch_score = batch_out[:,
+  1]).'