Add M1 submission (#23)

- Add M1 submission (5299d803260e3d33285ad2bd4c65112968098902)
- Fill reproduction block (match: scoring) (b49a69d61d68c60688563ac848aebb1ae1abad45)


Co-authored-by: Marius Bocanu <mbocanu@users.noreply.huggingface.co>

submissions/m1.yaml

@@ -0,0 +1,37 @@
+schema_version: 4
+system:
+  name: M1
+  slug: m1
+  description: 'M1, a voice deepfake detector (proprietary; weights not released; no paper). To keep the comparison fair,
+    we screened our training data against every Arena test set and removed every Arena test clip, and every other copy of
+    the same recording, from training (matched by source utterance). Where a set''s recording map was not available (ASVspoof
+    2021 LA and DF), we removed every recording of its evaluation speakers instead. We scored two training lengths of the
+    same model, and we submit the one that our own validation set preferred. Scoring: one window per clip from its first sample
+    (0.2–4.2 s for clips of 4.4 s or more, 0.2–2.2 s for 2.4–4.4 s, the whole clip if shorter; no padding or repetition).
+    Higher score = bona fide.'
+  code: https://huggingface.co/mbocanu/m1
+  checkpoint: https://huggingface.co/mbocanu/m1
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DFADD
+  revision: c578c836da3b522b27d3dd85f89309f1737e5d31
+  split: test
+scores:
+  eer_percent: 0.0
+  n_trials: 3755
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/mbocanu/m1/resolve/7b17027a944370cf784daf891b621bc5285e199f/.eval_results/SpeechAntiSpoofingBenchmarks/DFADD/scores.txt
+  scores_sha256: 558a5ecd759b9406659480fc5d96f57fffab7896b9e10d8ee54232d19da35b3d
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-10-27'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: scoring
+submitter:
+  hf_username: mbocanu
+  contact: imbocanu@gmail.com
+submitted_at: '2026-10-27'
+notes: Submitted on 23 of the 24 sets. We do not submit CFAD. CFAD labels its partially fake clips (real recordings with a
+  replaced segment) as fake at clip level. Our product scores speech segment by segment, and we do not think one clip label
+  is the right test for mixed clips.