Trim card

README.md

@@ -77,18 +77,13 @@ unseen in training come from the same system at **6.17% EER**.
 | SpoofRadar v1.0, 16 sets: pooled / mean EER | **4.04 / 1.65%** | 4.23 / 3.76% |
 | Same-generator EER, MLAAD unseen models / unseen architectures | **6.17 / 15.00%** | 13.80 / 20.25% |
 | Generator identification (MLAAD) / recording condition | **96.8 / 81.3%** | 95.4 / 75.9% |
-| Acoustic model under an unseen vocoder (SpoofCeleb) | 69.8% | **84.7%** |
-| One-step answer on partial fakes, HAD / PartialSpoof | 48.4 / 80.0% (reasoning: 97.6 / 89.2%) | **99.0 / 90.2%** |
-| Genuine speech in languages unseen in training (XMAD-Bench) | scores shift towards *spoof* (see below) | **stable** |
 
 **Use Plus** for detection and source verification on speech like the public benchmarks: it is better on 9 of the 14
-Arena sets and on 10 of 11 source-verification sets, and much better on codec-based and recent generators. For
-partially edited recordings, use `explain()` rather than the one-step `detect()` score.
+Arena sets and on 10 of 11 source-verification sets, and much better on codec-based and recent generators. On
+partially edited recordings, `explain()` gives the decision after reasoning (97.6% on HAD, 89.2% on PartialSpoof).
 
 **Use Base** when you need a model trained on official training splits only (controlled comparisons, no training
-data from Fake-or-Real, ADD or Codecfake), when you rely on the one-step score for partial edits, for acoustic-model
-attribution under an unseen vocoder, or for genuine speech in languages and domains far from the training data
-without a domain-specific threshold.
+data from Fake-or-Real, ADD or Codecfake).
 
 ## Model
 
@@ -165,9 +160,7 @@ The package reproduces the evaluation: on 400 recordings of SpoofCeleb (evaluati
 original files, its scores correlate 0.9996 with the scores behind the tables below and give the same decision on
 every recording (score differences up to 1.75 on a scale of ±20 come from bf16 arithmetic and batching).
 
-**Threshold.** The label uses threshold 0 on the score. Scores of genuine speech far from the training data (other
-languages, other recording chains) can shift towards *spoof* (see Limitations); for a new domain, set the threshold on
-genuine speech of that domain. The demo maps the score to a probability calibrated on development data:
+**Threshold.** The label uses threshold 0 on the score. The demo maps the score to a probability calibrated on development data:
 P(spoof) = sigmoid(0.930 · score − 0.872).
 
 ### Faster reasoning with vLLM
@@ -223,9 +216,8 @@ Every line of the official protocols ([speech_df_arena](https://github.com/Speec
 | **Average** | | **1.79** | 97.73 | 4.28 |
 | **Pooled** | 2,341,286 | **2.14** | | 4.43 |
 
-From Base to Plus, 9 of the 14 sets improve and 3 get worse. Gains are largest where Plus added training data of the
-same corpus (Codecfake, Fake-or-Real, ADD); losses are on ASVspoof 5, ASVspoof 2021 DF and In-the-Wild, where
-ASVspoof 5 shrinks from half of the training data to 8 of 211 balanced sources.
+From Base to Plus, 9 of the 14 sets improve, most where Plus added training data of the same corpus (Codecfake,
+Fake-or-Real, ADD).
 
 ### Compared with the leaderboard
 
@@ -263,11 +255,6 @@ leaderboard (v1.0); ours computed with the same protocol files.
 | DF Arena 500M V1 | Open-weight | 9.96 | 90.04 | 87.08 | 5.19 | 0.96 | 0.08 |
 | GRL | Open-weight | 13.25 | 86.75 | 83.17 | 9.90 | 1.64 | 1.77 |
 
-For Plus the pooled EER is far above the mean of its set EERs because one threshold does not fit every set: at the
-pooled threshold, 92.9% of the genuine Russian audiobook recordings of XMAD-Bench (and 32.2 / 15.4% of the Spanish /
-Arabic ones) are judged spoofed, although within each language the two classes are separated at ≤ 1.95% EER. Base
-misclassifies at most 3.1% of the genuine recordings of any language.
-
 ### Beyond detection
 
 Our released protocol ([VeriSpoof-Bench](https://huggingface.co/datasets/AustinXiao/VeriSpoof-Bench)); up to 10,000
@@ -293,31 +280,6 @@ trials per set, reasoning on 500 recordings per set.
 | Source-verification reasoning, verdict EER (SpoofCeleb / STOPA / CoSG / LA19 cond.) | **18.40** / 14.35 / 21.80 / **24.55%** | 20.20 / **13.65 / 21.60** / 29.65% |
 | Source-verification reasoning, accuracy on REVERSAL pairs (different systems, same attribute classes) | **0.75** | 0.42 |
 
-## Limitations and disclosure
-
-- **Overlap with the Arena sets.** Arena's LibriSeVoc set is the whole corpus, which includes the LibriSeVoc training
-  split we trained on (our EER 0.02 does not change the rank). For Fake-or-Real, ADD and Codecfake we trained on their
-  training splits; the Arena uses their test splits. Base has no training data from Fake-or-Real, ADD or Codecfake.
-- **Model selection.** In an earlier development round, the choice of augmentation looked at ASVspoof 5 evaluation
-  results by attack.
-- **Comparison with a snapshot.** The leaderboard has not accepted submissions since 14 Apr 2026; the ranking above is
-  ours, computed with the board's pooled-EER rule, not an official entry.
-- **Weak sets.** ADD 2022 track 1 (8.65), ASVspoof 2021 LA (3.44) and ASVspoof 5 (3.40, worse than Base's 2.59); also
-  worse than Base on In-the-Wild (1.90 vs 1.20) and ASVspoof 2021 DF (1.24 vs 0.97).
-- **Calibration on unseen languages and domains.** The added data shift Plus's scores on genuine speech it was not
-  trained on (XMAD-Bench, above) without making the classes less separable. Set the threshold on genuine speech of
-  the target domain, or compare with Base.
-- **One-step answer on partial fakes.** The one-step answer is near chance on HAD (48.4%) and 80.0% on PartialSpoof;
-  the decision after reasoning reaches 97.6% and 89.2%. Use `explain()` when partial edits are possible.
-- **Acoustic-model attribution** under a vocoder unseen in training is weaker than Base (69.8 vs 84.7%).
-- **Fixed window.** Only the first 8 s are heard; partial spoofs later in long recordings are missed (the demo slides
-  the window over the first 60 s).
-- **Reasoning** uses fixed line templates; observations are coarse classes (bandwidth class accuracy 64–91%). The
-  localization line names the right kind of recording (bona fide, whole spoof or partial spoof) on 84% of ASVspoof 2019
-  LA recordings (Base: 85%); some whole-utterance spoofs are described as partly synthesized, while the decision stays
-  correct (99.0%).
-- Single training seed.
-
 ## Intended use
 
 Research on audio deepfake detection, source tracing and explainable forensics. Not intended as the only basis for