WordSpoof is CC BY 4.0. The compilation and every synthetic clip are CC BY 4.0, so training, fine-tuning, evaluation and benchmarking are permitted for any purpose, commercial included; the only condition is attribution. The bona fide clips retain their upstream terms, which are at least as permissive: CC BY 4.0 for the Speech Commands subset, CC0 1.0 for the Common Voice subset. By requesting access you agree that you will attribute the dataset and its source corpora as set out in this card and in NOTICE.md; that you will honour the per-provider obligations in NOTICE.md, which travel with the data and include attribution and non-endorsement terms; and that you will not use any of it to impersonate any person, nor for fraud or deception.

Dataset Card for WordSpoof

A public corpus of standalone single-word bona fide and synthetic speech, for research on synthetic-voice detection. Most audio-deepfake corpora are sentence-length; this one is not, and detectors trained on sentence-length audio degrade sharply on it.

Dataset Details

Dataset Description

WordSpoof pairs real single-word utterances with synthetic ones generated by 10 modern text-to-speech and voice-cloning systems, so a detector — or a researcher auditing one — can measure what happens when the input shrinks from a sentence to a word. Both classes are English, drawn from single_word and yes_no response kinds, and shipped as 16 kHz mono PCM16 WAV with a per-clip JSON manifest.

Curated by: Mario Gómez-Martínez (Clearspeed) and James Gibson (Clearspeed)
Funded by: Clearspeed
Shared by: Clearspeed
Language(s) (NLP): English (en)
License: CC BY 4.0 over the compilation and every synthetic clip — commercial use permitted, attribution required. The bona fide clips retain their upstream CC BY 4.0 (Google Speech Commands v2) and CC0 1.0 (Mozilla Common Voice) terms, which are at least as permissive. Full breakdown, the legal reasoning, and per-provider obligations are in Licence and Legal Notices below — read that section before redistributing.

Dataset Sources

Repository: this HuggingFace dataset and a permanent archive copy at Zenodo, DOI 10.5281/zenodo.22763850. Both copies are byte-identical.
Paper: ICASSP 2027 submission titled "WordSpoof: a public dataset for synthetic-voice detection on single-word utterances", citing this dataset by its sha256. Citation details below.

Verify what you downloaded. SHA256SUMS.txt ships beside these files and covers all of them:

shasum -a 256 -c SHA256SUMS.txt

The published artifact is:

WordSpoof.tar
sha256  6c31dfc242279456d71b6c50e2e14322f6418437933bc5caf0b779219dc9921c
bytes   9,532,989,440
built   2026-09-13

Any paper reporting numbers on this corpus should cite that sha256, not the name — see Citation below for why.

Uses

Direct Use

Training, fine-tuning, evaluating, or benchmarking synthetic-voice detectors, for any purpose, commercial or not. The publisher adds no no-training term and no non-commercial term; the only condition is attribution. See Licence and Legal Notices for the exact per-subset terms, and NOTICE.md for the per-provider obligations that travel with the data.

The dataset is well suited to studying how detectors trained on sentence-length audio behave on single-word or short-utterance input, and to auditing content-blind shortcuts in existing detectors (see Bias, Risks, and Limitations, which documents one such shortcut measured on this corpus).

Out-of-Scope Use

Use without attribution. CC BY 4.0's one condition. The required form is in Licence and Legal Notices, and several generators add their own attribution and non-endorsement terms in NOTICE.md.
Impersonation. No clip may be used to impersonate a real person; every cloning reference voice is a consenting, published research speaker (VCTK), and every synthetic clip is labelled spoof.
Voice-cloning misuse of any kind, per the gating terms above.

Dataset Structure

WordSpoof/
  audio/<audio_id>.wav      one clip per file, 16 kHz mono PCM16
  manifest.jsonl            one JSON object per clip
  prompts.json              the prompt set
  release_meta.json         build provenance
  LICENSE
  NOTICE.md                 per-provider obligations and source attributions
  licenses/                 archived licence texts

Extract with tar -xf WordSpoof.tar, which yields a single WordSpoof/ directory.

release_meta.json carries both names: "release_name": "WordSpoof" and "build_id": "svd_short_public_v1". WordSpoof is what this dataset is called; the build id is the internal build the bytes came from, kept so the query_sql and source_table_versions in that file stay attributable to a real build.

Each manifest.jsonl row carries audio_id, audio_path, path_scope, dataset, split, label, lang, response_kind, prompt_id, prompt_text, prompts_revision_id, source_dataset, provider, provider_version, voice, voice_ref_source, take_idx, recipe_id, recipe_version, duration_s, sample_rate, codec, and a per-clip sha256.

Three of those carry more weight than their names suggest, and the guidance in this card depends on them:

recipe_id — the channel-conditioning recipe applied to the clip, and empty when none was. It is how you separate the 33% of spoof clips that ship as raw generator output from the 67% that were conditioned. Both Bias, Risks, and Limitations and the defect manifest turn on this distinction.
take_idx — which of the repeated generations for a given (voice, prompt) this clip is. Needed to treat (voice, prompt) as the sampling unit rather than the clip, which Recommendations asks you to do for kokoro.
path_scope — always tar_relative here, meaning audio_path resolves inside this archive.


clips	207,173
bona fide	138,555
spoof	68,618
spoof systems	10
language	English
audio	80.1 h · 16 kHz mono · PCM16 WAV
response kinds	`single_word`, `yes_no`
splits	train 168,375 · val 19,594 · test 19,204

Bona fide clips come from Google Speech Commands v2 (gsc_v2) and the single-word English segment of Mozilla Common Voice 7.0 (cv_singleword_7_0_en).

Spoof clips are generated by this project from 10 systems, a mix of preset-voice and zero-shot cloning:

generator	clips	generator	clips
`qwen_base`	15,488	`piper`	1,887
`cosyvoice2`	15,130	`kokoro`	1,863
`voxcpm2`	14,996	`qwen_cv`	1,203
`zonos`	14,283	`openai`	759
`chatterbox`	2,334	`magpie`	675

Total 68,618. The cells are very unequal — the four largest carry 87% of the spoof side, and magpie has 675 clips against qwen_base's 15,488. A pooled number over this corpus is dominated by four generators, so report per-generator results alongside any pooled figure.

Licence compatibility is a selection criterion for this set, not an afterthought. Every generator here is licensed such that its outputs can be redistributed and used commercially, and such that no obligation it imposes has to be passed on to you beyond the attribution and non-endorsement terms in NOTICE.md. A model whose terms reach its own outputs with a condition CC BY 4.0 §2(a)(5)(B) forbids us from imposing on you cannot appear in this corpus, however good it sounds.

For kokoro, the clip count is not a sample size — see Bias, Risks, and Limitations.

Dataset Creation

Curation Rationale

Most public audio-deepfake corpora are sentence-length. Deployed voice interfaces frequently are not — a spoken PIN digit, a "yes", a wake word. This corpus exists to measure detector behaviour on that shorter, differently structured input, and to do so with a public, redistributable artifact so published results can be reproduced independently.

Source Data

Cloning references come only from the VCTK corpus (CC BY 4.0, CSTR University of Edinburgh) — consenting, published research speakers. No private individual's voice was cloned. Every synthetic clip is labelled spoof.

Data Collection and Processing

Every spoof clip is scored with CTC ASR for structural defects — empty audio, runaway duration, or more content than was asked for.

Measured over all 68,618 spoof clips in this release, on two axes, because the first one structurally cannot see the second:

generator	shipped	generator-defect	quiet as shipped	either	either %
`qwen_cv`	1,203	115	2	115	9.56%
`cosyvoice2`	15,130	506	571	908	6.00%
`openai`	759	27	37	38	5.01%
`chatterbox`	2,334	62	1	63	2.70%
`voxcpm2`	14,996	9	71	80	0.53%
`qwen_base`	15,488	1	77	78	0.50%
`piper`	1,887	7	0	7	0.37%
`magpie`	675	1	0	1	0.15%
`zonos`	14,283	10	9	19	0.13%
`kokoro`	1,863	0	0	0	0.00%
total	68,618	738	768	1,309	1.91%

Axis 1, generator defects — 738 clips (1.08%). The three rules above, scored on raw generator output.

Axis 2, near-silence as shipped — 768 clips (1.12%). Peak amplitude below 0.05 in the delivered WAV. The audit cannot see these: its "empty" rule thresholds voiced frames relative to each clip's own maximum, so a uniformly quiet clip is never "empty". Only 197 clips appear on both axes.

Most near-silence is not the generator's fault, and the manifest says so per clip. Of the 768, only 197 had raw audio the audit also condemned; the other 571 left the generator clean and were attenuated downstream — 501 of them under the gscv2_noise_mp3_v1 conditioning recipe. They are listed anyway, because a near-inaudible clip is useless to you whichever stage made it quiet.

cosyvoice2 and qwen_cv are where the defects concentrate. Between them they carry 1,023 of the 1,309, on 24% of the spoof side. If you filter one cell, filter cosyvoice2. Three generators — piper, kokoro, magpie — are at or below 0.37%, and kokoro has none at all.

You can subtract them. A per-clip list ships beside the tarball:

WordSpoof_defects.csv     1,309 rows
WordSpoof_defects.README.md

Join it on audio_id against manifest.jsonl and drop the matches. What remains is 67,309 spoof clips and 205,864 rows in total, with all ten generators retained — bona fide clips are untouched. The basis column lets you subtract one axis rather than both.

The false-positive floor of the test itself. A defect rate means little without knowing how often the test is wrong. The same structural rule layer was run over 4,482 real human clips from the Speech Commands test split — the entire eligible pool, excluding the corpus's own _silence_ and _background_noise_ labels:

rejected: 1 of 4,482  =  0.022%

This is a property of the rule layer and of the bona fide corpus rather than of the spoof clips. One caveat stated precisely: the release ships Speech Commands' train/val split (105,829 clips) and this was measured on its test split, so it is a held-out estimate from the same corpus and recording conditions rather than a measurement of the shipped clips themselves.

Who are the source data producers?

Bona fide speech: Google (Speech Commands v2) and the Mozilla Common Voice project's contributors (single-word English segment, Common Voice 7.0). Cloning reference voices: consenting, published speakers from the VCTK corpus (CSTR, University of Edinburgh). Synthetic speech: 10 TTS/voice-cloning systems listed in Dataset Structure, run by the dataset curators.

Annotations

Not applicable in the crowd-annotation sense. Labels (bonafide / spoof) are assigned at generation time by construction, not by human annotators. Structural-defect verdicts (see Data Collection and Processing) are produced by an automated CTC-ASR-based audit, not by manual annotation.

Personal and Sensitive Information

The bona fide clips are not filtered and carry no defect list. A glitchy synthetic clip is a generation failure; a noisy human recording is a real recording, which is what a bona fide corpus is for.

No private individual's voice was cloned. Cloning references come only from consenting, published research speakers (VCTK). Every synthetic clip is labelled spoof, satisfying disclosure obligations that attach to cloning use (see Licence and Legal Notices). We are not aware of any personal, sensitive, or private information in the bona fide source corpora beyond what their own publishers (Google, Mozilla) have already disclosed and licensed for redistribution.

Bias, Risks, and Limitations

The corpus carries a measured low-level shortcut. A classifier can separate spoof from bona fide using content-blind features alone — no ASR, no transcript, no vocabulary. Measured on the clips in this tarball, 400 per (provider × conditioning) cell sampled deterministically by audio_id hash, 5-fold cross-validated AUC:

subset	all features	framing only	acoustics only
`single_word`	0.9419	0.8510	0.8821
`yes_no`	0.9812	0.8631	0.9401

"Framing" is duration, leading and trailing silence, and speech span; "acoustics" is everything else — energy, spectral tilt, edge behaviour. Both groups separate the classes well on their own, so this is not a single fixable artefact.

yes_no is the worse cell, and the reason is corpus composition, not the generators. Its bona fide side is a single corpus of 4,868 clips from one recording channel (cv_singleword_7_0_en), where single_word draws on two corpora and 133,687 clips (gsc_v2 105,829 + cv_singleword_7_0_en 27,858). A narrower real-speech distribution is easier to separate from anything. Per spoof cell, against pooled bona fide, all 20 cells of each response kind land between 0.9155 and 1.0000 — so no single generator is carrying this.

Clip duration is a provider fingerprint on the unconditioned clips, and for one generator it is exact. Fraction of clips whose duration is an exact multiple of 0.16 s — the frame period of a 6.25 Hz acoustic token rate — measured on the shipped WAVs:

generator	unconditioned	conditioned
`voxcpm2`	100.0% (5,004)	1.7% (9,992)
`qwen_base`	49.0% (5,162)	1.7% (10,326)
`qwen_cv`	47.1% (397)	3.3% (806)
`chatterbox`	23.9% (778)	2.5% (1,556)
every other generator	≤ 6.3%	≤ 2.7%

A duration is 16 kHz sample-quantised, so 0.16 s is 2,560 samples and the chance rate is 1/2,560 = 0.04%. Against that, voxcpm2's unconditioned clips are a perfect rule: duration mod 0.16 s == 0 identifies them without listening. qwen_base and qwen_cv sit near 50% because their token rate is twice as fast, so half their clips coincide with the same grid.

Conditioning destroys the pattern — every conditioned cell falls to the 1-3% range — because the noise recipes repad the clip. So this affects the 22,863 spoof clips (33%) that ship unconditioned. Anything you train should be checked against a duration-only baseline, and if you need the effect gone, drop the unconditioned spoof rows (recipe_id is empty for them in manifest.jsonl).

Near-silence is unevenly distributed across generators, which is a label rule of the same kind. 768 spoof clips ship with peak amplitude below 0.05, and the rate per generator runs from 4.87% (openai) and 3.77% (cosyvoice2) down to 0.00% (piper, kokoro, magpie). Most of it is introduced by conditioning rather than by the generator — 571 of the 768 had clean raw audio — but a detector cannot tell the difference and neither can a downstream metric. Every one of those clips is enumerated in WordSpoof_defects.csv so you can subtract them; see Data Collection and Processing.

kokoro rows are not three independent samples. kokoro was generated with three takes per (voice, prompt), as the other generators were. For kokoro those three are not three independent samples, and the reason is architectural rather than a property of any particular batch. Read from the kokoro 0.9.4 package: KModel.forward(phonemes, ref_s, speed) is its entire inference surface — no seed, no temperature, no sampling argument anywhere in it. Duration is a deterministic integer frame count, round(sigmoid(duration).sum(-1) / speed), and the style vector is a fixed voicepack lookup. There is no way to ask kokoro for a different reading of a prompt. So sibling takes share the same words, voice, prosody, F0 contour, phoneme alignment and generated duration by construction; what differs is the decoder's random harmonic phase, which eval() does not disable. A re-excitation of one performance, not a second performance. The other generators here expose a seed or a temperature, so this is specific to kokoro rather than a property of the take protocol.

Two consequences for anything you compute on the kokoro cell:

Its effective sample size is about a third of its clip count. The 1,863 kokoro clips carry about 621 independent utterances — exactly 3.0 clips per unit. Treat (voice, prompt) as the unit, not the clip; a standard error computed per clip is understated by roughly √3.
Near-duplicates span the train/test split. The split is assigned per clip and the three takes are distinct clips, so they can land on different sides. A detector trained on the train split has seen a re-excitation of some test utterances. This is not an identity leak and the audio is not identical — but a kokoro test number is optimistic by an unmeasured amount.

Whether this is a defect depends on what you are measuring. A spoof detector keys largely on vocoder artefacts, and a re-randomised excitation is arguably a fresh draw from that distribution. For prosody, duration or content-level analysis it is not. Only kokoro is affected — magpie's and openai's handful of duration-matching cross-split pairs were checked at byte level and are coincidence, 93-99% of their samples differ. This is documented rather than filtered so you can decide for yourself; a future build caps kokoro at one take per combo.

OpenAI is the only commercial API among the ten generators. Detector results measured here may therefore not transfer to the strongest commercial cloning services, and work using this corpus should state that limit.

The structural defect test does not catch identity miss — a fluent, correctly-timed clip that says the wrong words. That is a separate axis and it is not measured in this release.

Recommendations

A duration-only or acoustics-only baseline is mandatory for any result claimed on this corpus. The shortcut above is measured, not estimated, and it is reported rather than filtered out, so its size is known to you before you build on the data. Acoustics alone reach 0.88 and 0.94 AUC; a model that does not beat that has learned nothing about synthesis.
Check any result against the duration grid, or drop the unconditioned spoof rows. voxcpm2's unconditioned clips are identifiable from their duration alone, exactly.
Report single_word and yes_no separately. They are 0.04 apart on the shortcut measurement, which is larger than most reported improvements, and pooling them hides it.
Use the defect manifest (WordSpoof_defects.csv) if you need the clean spoof subset — and decide which axis you care about. Generator defects and near-silence are separate columns because 571 of the 768 near-silent clips were attenuated by conditioning rather than mis-generated.
Report per-generator figures alongside any pooled one — the generator cells are very unequal (87% of the spoof side from four generators), and defect rates range from 0.00% to 9.56% across them.
Treat kokoro's (voice, prompt) pair as the sampling unit, not the clip, when computing standard errors or designing a train/test split.

Licence and Legal Notices

This release is CC BY 4.0, with one carve-out that is more permissive, not less. The provider and source_dataset fields in manifest.jsonl tell you which clips are which.

what	licence	clips
the compilation — manifest, prompt set, split assignment, selection and arrangement	CC BY 4.0	—
every synthetic clip	CC BY 4.0	68,618
bona fide from Google Speech Commands v2 (`gsc_v2`)	CC BY 4.0	105,829
bona fide from Common Voice single-word EN (`cv_singleword_7_0_en`)	CC0 1.0 (public domain)	32,726

Licence texts are in licenses/.

Identify your copy by sha256. This release is 6c31dfc242279456d71b6c50e2e14322f6418437933bc5caf0b779219dc9921c, and the licence text that governs it ships inside the tarball as LICENSE.

You do not have to take our word for the bona fide half. The check is outside this release: Google publishes Speech Commands v2 under CC BY 4.0 and Mozilla publishes the Common Voice single-word English segment under CC0 1.0.

Why a single permissive licence is available here. CC BY 4.0 §2(a)(5)(B) bars a licensee from imposing additional or different terms on licensed material. So a generator whose own licence reaches its outputs with a condition that would bind you — a naming requirement on models you train, or a non-commercial limit — cannot be in this corpus, because we could neither pass that condition on to you nor honestly stay silent about it. Every generator in Dataset Structure clears that bar on its own terms.

Why Common Voice stays CC0. It is public domain, and nothing here claims rights over it. CC BY 4.0's attribution condition does not attach to those clips, though the courtesy credit below is given regardless.

Attribution, required for the CC BY 4.0 material:

Warden, P. Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition, Google, 2018 — for the gsc_v2 clips.
Mozilla Common Voice 7.0, single-word English segment — for the cv_singleword_7_0_en clips.
This release, cited as below, for the compilation and the synthetic clips.

Training and commercial use are both expressly permitted. You may use these clips to train, fine-tune, evaluate and benchmark models for any purpose, commercial or not, and you may distribute what you build. The publisher adds no no-training term and no non-commercial term. The only condition is attribution, in the form given above, plus the per-provider obligations below — which are attribution and non-endorsement terms, not use restrictions.

There is no research-versus-product distinction to police here: a checkpoint trained on this corpus may become a product checkpoint.

Obligations that travel with the data, listed per provider in NOTICE.md. In summary:

All spoof clips are AI-generated and the generating system is named per clip. Several generators require this disclosure; the spoof label and the provider field satisfy it.
Magpie is licensed under the NVIDIA Open Model License, which requires attribution on redistribution. NVIDIA claims no ownership of outputs.
LibriTTS (OpenSLR 60, CC BY 4.0) is attributed for the Piper en_US-libritts-high voice — Zen, H. et al., LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech.
Kokoro's model card attributes its training audio to the Koniwa corpus (CC BY 3.0) and the SIWIS corpus (CC BY 4.0).
Chatterbox clips carry Resemble AI's Perth perceptual watermark, on by default. This is a methodology warning rather than a licence one: a detector trained on that partition may learn the watermark instead of genuine synthesis artefacts. No trained detector checkpoints are released. That is a scope decision about this release, not a licence restriction: under CC BY 4.0 you may train on this corpus and distribute the result, subject to attribution.

Citation

BibTeX:

@misc{gomez2026svdshort,
  author       = {Gómez-Martínez, Mario and Gibson, James},
  title        = {{WordSpoof}: a public corpus of single-word bona fide and
                  synthetic speech},
  year         = {2026},
  publisher    = {Zenodo},
  version      = {v1 (artifact 6c31dfc24227)},
  doi          = {10.5281/zenodo.22763850},
  howpublished = {Zenodo, \url{https://doi.org/10.5281/zenodo.22763850}},
  note         = {sha256 6c31dfc242279456d71b6c50e2e14322f6418437933bc5caf0b779219dc9921c}
}

@misc, deliberately, not @dataset. Several IEEE and ACM styles — including IEEEbib.bst, which ICASSP uses — have no @dataset entry type and silently drop publisher, version and doi from the rendered reference. The howpublished field is there so the DOI still renders under those styles.

APA:

Gómez-Martínez, M., & Gibson, J. (2026). WordSpoof: a public corpus of single-word bona fide and synthetic speech (Version v1, artifact 6c31dfc24227) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22763850

Glossary

Structural defect — a spoof clip whose audio is empty, runs away in duration, or contains more content than the prompt asked for. Measured with CTC ASR on raw generator output. Distinct from near-silence, which is measured on the shipped bytes and which this test cannot detect; both ship in WordSpoof_defects.csv. See Data Collection and Processing.
Identity miss — a fluent, correctly-timed synthetic clip that says the wrong words. Not caught by the structural-defect test; not measured in this release.
Low-level shortcut — separability between spoof and bona fide using content-blind features alone (no ASR, no transcript). Measured on this corpus and disclosed in Bias, Risks, and Limitations.
response_kind — the manifest field distinguishing single_word clips from yes_no clips; report results per kind, not pooled.
build_id — the internal build identifier recorded in release_meta.json, distinct from release_name (WordSpoof); see Dataset Structure.

More Information

The paper this dataset supports is a submission to ICASSP 2027, titled "WordSpoof: a public dataset for synthetic-voice detection on single-word utterances"; citation details will be added to this card once accepted. For questions about a specific provider's licence terms, consult NOTICE.md inside the archive first — it is the per-provider source of truth and is machine-checkable against manifest.jsonl's provider field.

Dataset Card Authors

Mario Gómez-Martínez (Clearspeed) and James Gibson (Clearspeed)

Dataset Card Contact

mario.gomez@clearspeed.com

Downloads last month: 21

Total file size:

9.53 GB