Fix: exclude seedtts_testset (real reference prompts mislabeled spoof); 4548->3948

README.md

@@ -28,11 +28,11 @@ synthetic / deepfake voice detection.
 
 ## Overview
 
-SONAR is a balanced binary-classification benchmark: **bonafide** (genuine human speech)
-vs. **spoof** (AI-synthesized speech). The spoof side aggregates clips from **nine modern
+SONAR is a binary-classification benchmark: **bonafide** (genuine human speech)
+vs. **spoof** (AI-synthesized speech). The spoof side aggregates clips from **eight modern
 speech-synthesis systems**, deliberately spanning architectures and providers so that a
-detector cannot win by memorising one generator's artifacts. The bonafide side is an equal
-number of real clips sampled from the **LibriTTS** clean-test set. Original code and data:
+detector cannot win by memorising one generator's artifacts. The bonafide side is real
+clips sampled from the **LibriTTS** clean-test set. Original code and data:
 https://github.com/Jessegator/SONAR.
 
 | Side | Source | n |
@@ -40,7 +40,6 @@ https://github.com/Jessegator/SONAR.
 | bonafide | LibriTTS clean-test | 2274 |
 | spoof | OpenAI TTS | 600 |
 | spoof | xTTS | 600 |
-| spoof | Seed-TTS (en 300 + zh 300) | 600 |
 | spoof | FlashSpeech | 118 |
 | spoof | VoiceBox | 104 |
 | spoof | AudioGen | 100 |
@@ -48,6 +47,12 @@ https://github.com/Jessegator/SONAR.
 | spoof | NaturalSpeech 3 | 32 |
 | spoof | PromptTTS 2 | 25 |
 
+> **Seed-TTS excluded (2026-06-07).** The upstream `seedtts_testset/` directory holds the
+> SeedTTS *reference prompts* — real Common Voice audio (`common_voice_en_<prompt>-<target>.wav`),
+> not synthesized speech (the actual SeedTTS-generated outputs are not redistributed on disk).
+> Labeling those 600 real clips as spoof was incorrect, so they are dropped from this package.
+> This matches the SONAR authors' own released eval set (3948 clips).
+
 ## License & redistribution
 
 This **packaging** is released under the **Creative Commons Attribution-NonCommercial 4.0
@@ -86,16 +91,16 @@ print(ds[0])
 
 | Stat | Value |
 |------|-------|
-| Total trials | 4548 |
+| Total trials | 3948 |
 | Bonafide | 2274 |
-| Spoof | 2274 |
+| Spoof | 1674 |
 
 ## Source provenance
 
 - Paper: https://arxiv.org/abs/2410.04324
 - Code / data: https://github.com/Jessegator/SONAR
-- Labels derived from the source directory layout (`real_samples/` = bonafide; the nine
-  synthesis-system directories = spoof).
+- Labels derived from the source directory layout (`real_samples/` = bonafide; the eight
+  synthesis-system directories = spoof; `seedtts_testset/` excluded — real reference prompts).
 
 ## Evaluation
 

build_parquet.py

@@ -52,14 +52,20 @@ except AttributeError:  # older datasets
 REPO_ROOT = Path(__file__).resolve().parent
 SRC_ROOT = Path("/home/kirill/mnt/users_4tb/datasets/sonar")
 REAL_DIR = "real_samples"  # bonafide; every other top-level dir is spoof
+# EXCLUDED dirs: seedtts_testset holds the SeedTTS *reference prompts* (real Common
+# Voice audio, common_voice_en_<prompt>-<target>.wav), NOT synthesized speech — the
+# actual SeedTTS-generated outputs are not on disk. Labeling them spoof mislabels 600
+# real clips (Spectra-AASIST scored 99.83% bona fide; full-set EER 24.85% vs 0.48%
+# without them). Matches the SONAR authors' own 3948-file eval. (2026-06-07)
+EXCLUDE_DIRS = {"seedtts_testset"}
 PARQUET_DIR = REPO_ROOT / "data"
 STAGING_DIR = REPO_ROOT / "_clean_flac"
 NUM_SHARDS = 2
-EXPECTED_ROWS = 4548
+EXPECTED_ROWS = 3948
 EXPECTED_BONAFIDE = 2274
-EXPECTED_SPOOF = 2274
+EXPECTED_SPOOF = 1674
 TARGET_SR = 16000
-PROBE_SAMPLE = 4548  # whole set
+PROBE_SAMPLE = 3948  # whole set (seedtts_testset excluded)
 WORKERS = int(os.environ.get("SONAR_BUILD_WORKERS", os.cpu_count() or 4))
 
 FEATURES = Features(
@@ -84,6 +90,8 @@ def build_catalogue():
     for path in SRC_ROOT.rglob("*.wav"):
         relpath = path.relative_to(SRC_ROOT).as_posix()
         system = relpath.split("/", 1)[0]
+        if system in EXCLUDE_DIRS:
+            continue
         label = "bonafide" if system == REAL_DIR else "spoof"
         language = "zh" if "/zh/" in f"/{relpath}" else "en"
         records.append(

data/labels.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9edb7158cc6241bed48edcff96d10dc85f3b977b911dc4cd26ae3144584cb7c6
-size 35013
+oid sha256:e05a59ad837b922948aac0f3a8ef70a752a247bf652e275e0a5994618afc8fcf
+size 28035

data/test-00000-of-00002.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a1b7ebe9290425834bf8fcd0e471806247c68fa0a36979372e6a7d7a0f08f750
-size 288064744
+oid sha256:e398e67c65b0fe18ef6ac45a9d0f4f8c4685d62c2cab6b53f7d2ce381e22ef2d
+size 233448507

data/test-00001-of-00002.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fe7259b670680394ca2e1d4477ba41e2ac0b6bb940c61ef8014a1e1487002e32
-size 290017252
+oid sha256:54c9b77e96202ca9ea73c309e26d8f7693a305619fcf0143ec86581e4aa73a72
+size 293936334

eval.yaml

@@ -1,10 +1,11 @@
 name: SONAR
 description: >
-  SONAR synthetic-audio detection benchmark (arXiv 2410.04324). Balanced binary
-  classification: 2274 bonafide (LibriTTS clean-test) vs. 2274 spoof aggregated
-  from nine speech-synthesis systems (OpenAI TTS, xTTS, Seed-TTS, FlashSpeech,
-  VoiceBox, AudioGen, VALL-E, NaturalSpeech 3, PromptTTS 2). EER computed over all
-  4548 utterances.
+  SONAR synthetic-audio detection benchmark (arXiv 2410.04324). Binary
+  classification: 2274 bonafide (LibriTTS clean-test) vs. 1674 spoof aggregated
+  from eight speech-synthesis systems (OpenAI TTS, xTTS, FlashSpeech, VoiceBox,
+  AudioGen, VALL-E, NaturalSpeech 3, PromptTTS 2). EER computed over all 3948
+  utterances. (Seed-TTS excluded: the on-disk seedtts_testset clips are the real
+  Common Voice reference prompts, not synthesized speech.)
 evaluation_framework: inspect-ai
 
 tasks: