vouch-bench: synthetic-speech detection under messaging-app conditions

What this is. A small benchmark of real and synthetic speech, shaped like messaging-app voice notes: 2-10 s clips, Opus at 16 and 24 kbps, a phone-microphone simulation, and English plus Japanese. Every clip comes in four renderings (clean, Opus 16k, Opus 24k, phone mic). The folds test whether a detector generalises to a generator it has never seen.

What it can measure.

• Detection with all generators seen, under each messaging condition.
• Leave-one-generator-out generalisation, separately within the voice-conversion (VC) lineage and across lineages.
• How a detector trained on older generators treats a modern neural TTS (Kokoro-82M) that sits in a test-only partition and must never be used for training, validation, normalisation statistics or model selection.

What it cannot measure.

• Unseen VC lineage is at or below chance for every detector we tried. With no VC in training, the reference detector reaches AUC 0.503 on the research benchmark and 0.351 on this release subset (decisions invert). The benchmark exposes the failure; it cannot tell you how to fix it.
• Zero-shot voice cloning, commercial TTS APIs and phone-call audio are not included. Results do not transfer to them.
• Japanese results are exploratory. FLEURS has no speaker ids, so speaker-disjointness relies on six coarse (split, gender) groups, and the Japanese test side is female only.
• The data is small (1,000 base clips), so treat differences of a few points of EER as noise.

Composition

4,000 clips = 1,000 base clips x 4 messaging profiles. Base hours: 2.02.

label	generator	lang	base clips	clips (x4 profiles)
fake	espeak_ng	en	100	400
fake	espeak_ng	ja	100	400
fake	freevc	en	100	400
fake	freevc	ja	100	400
fake	knn_vc	en	100	400
fake	knn_vc	ja	100	400
fake	kokoro_82m	en	100	400
fake	kokoro_82m	ja	100	400
real	-	en	100	400
real	-	ja	100	400

test_only = true marks the 800 Kokoro-82M clips. They are a hold-out. Do not train, validate, tune or select on them.

macOS system-voice clips from the research benchmark are excluded from this release because their redistribution terms have not been cleared. The "say unseen" condition of the research benchmark therefore cannot be evaluated here.

Folds (folds/*.json)

Speakers and synthetic voices never cross between train and test. VC fakes take the identity of their target speaker, so they follow that speaker's side. All four renderings of a base clip stay on the same side.

fold	meaning	train clips	test clips
speaker	All generators seen	2164	1036
logo-espeak_ng	espeak-ng unseen (= TTS lineage unseen)	1588	1052
logo-freevc	FreeVC unseen (kNN-VC in train)	1868	532
logo-knn_vc	kNN-VC unseen (FreeVC in train)	1868	532
family-tts	TTS lineage unseen	1588	1052
family-vc	VC lineage unseen (both VCs held out)	1348	812

Only one TTS generator (espeak-ng) is trainable in this release, so family-tts and logo-espeak_ng hold out the same generator and have identical test sets. folds/holdout.json lists the test-only Kokoro clips.

Usage

from datasets import load_dataset
ds = load_dataset("NagaYu/vouch-bench", split="train")   # all rows; use folds/*.json to split

Or with pandas: pd.read_parquet("manifest.parquet"). The file_name column is relative to the repository root. For a fold, select clip_ids from folds/<fold>.json (train / test). The test-only hold-out is always test_only == True.

Key columns: label (real/fake), lang, speaker_id, generator_id (null for real), generator_family (tts/vc), codec_profile, duration, source_id, source_clip_id, paired_clip_id, test_only, license, sha256, gen_params.

Reference results

Benchmark tables (research setting, which also includes macOS-say clips in training) and the release-checkpoint table are in the model card: https://huggingface.co/NagaYu/vouch-hubert-lr.

Licences

There is no single licence. See LICENSES.md for the component table and full attribution: LibriSpeech, FLEURS and VCTK are CC BY 4.0; kNN-VC and FreeVC are MIT; Kokoro-82M is Apache-2.0; eSpeak NG is GPL software whose audio output the GPL does not cover. Each manifest row records its own source, generator and licence.

How it was made

See REPRODUCE.md.

Contact

Portfolio: https://nagayu-portfolio.nagayu.workers.dev/en