WordSpoof is double-licensed. The compilation and every synthetic clip are CC BY-NC 4.0. The bona fide clips retain their upstream terms and are more permissive: CC BY 4.0 for the Speech Commands subset, CC0 1.0 for the Common Voice subset. See the Licence and Legal Notices section of this card, which supersedes the LICENSE file inside the tarball on this point. By requesting access you agree that you will use the CC BY-NC material for non-commercial purposes only, including academic research and education; that you will attribute the dataset and its source corpora as set out in this card and in NOTICE.md; that you will not use the CC BY-NC material to build or improve any commercial product or service, and will not allow a model trained on it to become a commercial model; and that you will not use any of it to impersonate any person. Several spoof partitions carry further obligations from the model that generated them, listed per provider in NOTICE.md, and those travel with the data.

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 11 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 (Clearspeed) and Jimmy Gibson (Clearspeed)
Funded by: Clearspeed
Shared by: Clearspeed
Language(s) (NLP): English (en)
License: Double-licensed. CC BY-NC 4.0 covers the compilation and every synthetic clip; the bona fide clips retain their upstream CC BY 4.0 (Google Speech Commands v2) and CC0 1.0 (Mozilla Common Voice) terms. Full breakdown, the legal reasoning, and per-provider obligations are in Licence and Legal Notices below — read that section before redistributing or fine-tuning on any subset in isolation.

Dataset Sources

Repository: this HuggingFace dataset and a permanent archive copy at Zenodo, DOI 10.5281/zenodo.22655557. 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  a83249a373d2b5ced967a6e3f5289a21dd92de0d83e30afe1ca8f9535aba5dee
bytes   9,771,991,040
built   2026-09-01

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 non-commercial purposes, including academic research and education. The licence restricts commercial use, not training itself, and the publisher adds no additional no-training term. See Licence and Legal Notices for the exact per-subset terms and the reasoning behind them.

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

Any commercial exploitation of the CC BY-NC–licensed material (the compilation and all synthetic clips). This includes shipping a model trained on this corpus as, or inside, a commercial product — the restriction is retroactive: a research checkpoint must not become a product checkpoint.
Distributing trained detector checkpoints. Not done by this release and not licensed for you to do either — see Licence and Legal Notices.
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
  build_skips.jsonl         clips the build declined, with reasons
  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, split, label, lang, response_kind, prompt_id, prompt_text, source_dataset, provider, provider_version, voice, voice_ref_source, duration_s, sample_rate, codec, and a per-clip sha256.


clips	208,734
bona fide	138,555
spoof	70,179
spoof systems	11
language	English
audio	82.2 h · 16 kHz mono · PCM16 WAV
response kinds	`single_word`, `yes_no`
splits	train 169,641 · val 19,726 · test 19,367

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 11 systems, a mix of preset-voice and zero-shot cloning:

generator	clips	generator	clips
`qwen_base`	15,488	`piper`	1,887
`indextts2`	15,480	`kokoro`	1,863
`cosyvoice2`	15,130	`qwen_cv`	1,200
`zonos`	14,283	`orpheus`	1,080
`chatterbox`	2,334	`openai`	759
		`magpie`	675

Total 70,179. The cells are very unequal — the four largest carry 86% 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.

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 was scored with CTC ASR for structural defects — empty audio, runaway duration, or more content than was asked for. Measured 2026-09-04 over all 70,179 shipped spoof clips:

generator	shipped	defective	rate
`indextts2`	15,480	2,017	13.0%
`qwen_cv`	1,200	112	9.3%
`openai`	759	27	3.6%
`cosyvoice2`	15,130	506	3.3%
`chatterbox`	2,334	62	2.7%
`orpheus`	1,080	9	0.8%
`piper`	1,887	7	0.4%
`magpie`	675	1	0.1%
`zonos`	14,283	10	0.1%
`qwen_base`	15,488	1	0.0%
`kokoro`	1,863	0	0.0%
total	70,179	2,752	3.92%

By kind: runaway 1,359, runaway+excess_content 903, excess_content 456, empty 34.

indextts2 carries 2,017 of the 2,752 — 73% of all defects — on 22% of the spoof side. If you filter one cell, filter that one. Six generators are at or below 1%, and kokoro has none.

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

WordSpoof_defects.csv     2,752 rows
WordSpoof_defects.README.md

Join it on audio_id against manifest.jsonl and drop the matches. What remains is 67,427 spoof clips and 205,982 rows in total, with all 11 generators retained — bona fide clips are untouched, since this is a generator-defect list.

The same test rejects almost no real speech. 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%

So the 3.92% above is not an artefact of a trigger-happy detector. 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: 11 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, unfixed low-level shortcut. A classifier can separate spoof from bona fide using content-blind features alone — no ASR, no transcript, no vocabulary:

subset	pooled boosted AUC	acoustic features only
`single_word`	0.9377	0.8824
`yes_no`	0.9775	0.9406

yes_no is the worse cell because its bona fide side is a single corpus of 4,868 clips — one recording channel — where single_word draws on two corpora and 133,687 clips. It is not a generator artefact: all 22 spoof cells fall between 0.9671 and 1.0000.

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 eleven 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, given the low-level shortcut above. The shortcut is measured and reported rather than filtered out, so the size of it is known to you before you build on the data.
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 report per-generator figures alongside any pooled one — the generator cells are very unequal (86% of the spoof side from four generators).
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 DOUBLE-licensed. Which licence applies depends on which clips you use, and the provider and source_dataset fields in manifest.jsonl tell you exactly which is which.

what	licence	clips
the compilation — manifest, prompt set, split assignment, selection and arrangement	CC BY-NC 4.0	—
every synthetic clip	CC BY-NC 4.0	70,179
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	32,726

Licence texts are in licenses/.

You do not have to take our word for any of this. 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. Those upstream terms are what govern those clips, and §2(a)(5)(B) is why they cannot be narrowed here.

The tarball's own LICENSE and NOTICE.md state this same position. If you hold a copy whose LICENSE claims a single CC BY-NC 4.0 over every clip, it is an earlier build — check the sha256.

Why the bona fide clips keep their own terms. Speech Commands v2 is CC BY 4.0, and CC BY 4.0 §2(a)(5)(B) bars a licensee from offering or imposing additional or different terms that restrict exercise of the licensed rights. Non-commercial is such a restriction, so those clips cannot be relicensed NC. Common Voice is CC0, which has no term to breach — but public-domain material cannot have NC rights claimed over it either. So the bona fide side is more permissive than the release as a whole, and nothing here narrows it.

Attribution, required for the CC BY-NC and CC BY parts:

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 is expressly permitted. You may use these clips to train, fine-tune, evaluate and benchmark models for non-commercial purposes, including academic research and education. The CC BY-NC part restricts commercial use; it does not restrict training, and the publisher adds no no-training term of its own.

What is restricted is commercial exploitation of the CC BY-NC material — and that restriction turns on what a trained model is later used for, so a research checkpoint must not become a product checkpoint. Note that a model trained on the whole corpus is subject to the NC term even though part of its input was CC BY or CC0.

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.
bilibili does not endorse, sponsor or certify this work (IndexTTS-2 Model Use Agreement §4.1(a)); the Agreement ships in licenses/, in English and Chinese, and its terms flow down to you.
Orpheus carries "Built with Llama" and the Meta Llama 3.2 community licence notice.
Magpie is licensed under the NVIDIA Open Model License.
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.

No trained detector checkpoints are released. Distributing a model trained on this corpus would exit OpenAI's classifier exception and breach two providers' non-commercial terms.

Citation

BibTeX:

@misc{gomez2026svdshort,
  author       = {Gomez, Mario and Gibson, Jimmy},
  title        = {{WordSpoof}: a public corpus of single-word bona fide and
                  synthetic speech},
  year         = {2026},
  publisher    = {Zenodo},
  version      = {v1 (artifact a83249a373d2)},
  doi          = {10.5281/zenodo.22655557},
  howpublished = {Zenodo, \url{https://doi.org/10.5281/zenodo.22655557}},
  note         = {sha256 a83249a373d2b5ced967a6e3f5289a21dd92de0d83e30afe1ca8f9535aba5dee}
}

@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:

Gomez, M., & Gibson, J. (2026). WordSpoof: a public corpus of single-word bona fide and synthetic speech (Version v1, artifact a83249a373d2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22655557

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; 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 acoustic features alone (no ASR, no transcript). Measured 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 (Clearspeed)

Dataset Card Contact

mario.gomez@clearspeed.com

Downloads last month: -

Total file size:

9.77 GB