VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation
Paper • 2101.00390 • Published • 1
Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is derived from public European Parliament plenary footage via the CC0 VoxPopuli segment index. Access is gated to record use and ensure attribution. By requesting access you agree to (1) attribute "© European Union – Source: EP" in any use or redistribution, and (2) use the data for research consistent with the European Parliament's audiovisual reuse terms.
Log in or Sign Up to review the conditions and access this dataset content.
Ellipsis — European Parliament Mouth-ROI (multilingual VSR)
Mouth-region-of-interest (ROI) clips + verbatim transcripts for visual speech recognition (VSR / lip-reading), derived from European Parliament plenary debates (2019–2020) through the CC0 VoxPopuli segment index.
- 57,735 clips across 16 languages, in 58 WebDataset shards (~106 GB).
- Each clip is the speaker's mouth ROI only — a
[T, 96, 96]grayscale tensor (Auto-AVSR convention) — paired with the VoxPopuli verbatim transcript. No full frames, no faces beyond the mouth crop, no audio, no identity labels are stored (data-minimisation by design).
Contents
WebDataset tar shards; each sample is three members sharing one key:
| member | content |
|---|---|
<key>.roi.npy |
[T, 96, 96] uint8 grayscale mouth ROI (np.load-able) |
<key>.txt |
UTF-8 verbatim transcript (the label) |
<key>.json |
metadata: language, session_key, audio_id, prompt_id, corpus_tier, frame count |
dataset-manifest.jsonl carries the content-hash lineage of every shard.
Statistics
By language
| lang | clips | lang | clips | lang | clips | lang | clips |
|---|---|---|---|---|---|---|---|
| en | 17,782 | de | 9,516 | es | 7,189 | fr | 6,295 |
| it | 3,262 | pl | 2,815 | cs | 2,253 | ro | 2,082 |
| nl | 1,968 | hr | 1,474 | hu | 981 | sk | 781 |
| fi | 666 | sl | 519 | lt | 125 | et | 27 |
By year: 2019 → 41,528 · 2020 → 16,207. Sittings: 82 (49 in 2019, 33 in 2020).
How it was built
Segment spans + verbatim text come from VoxPopuli (CC0). The floor-track video is fetched per debate from the EP VOD, the Original (floor) audio track is verified against the manifest (never an interpretation booth), a face is detected + tracked (MediaPipe 478-landmark), the active speaker is confirmed (Light-ASD), and the mouth is warped/cropped to a fixed 96×96 grayscale ROI. A per-session audio/video offset (human-judged, "zero it by ear") is applied so lips and voice are aligned — 2019 sittings carry offsets up to ~1.9 s at the source; 2020 sittings are aligned at 0.
Intended use & tiers
- Multilingual VSR pretraining — all 16 languages.
- Italian verbatim fine-tuning — the
itsubset (3,262 clips) carries gold VoxPopuli transcripts; filterlanguage == "it"in the.json. Non-Italian clips are tagged as pretraining-tier.
Licensing & attribution
Released under CC0-1.0 following VoxPopuli's CC0 licensing of the derived plenary data. Attribution is required: "© European Union – Source: EP", consistent with the European Parliament's audiovisual reuse terms. VoxPopuli: Wang et al., 2021 (arXiv:2101.00390).
Ethics & privacy
The subjects are public figures (Members of the European Parliament) speaking in public plenary proceedings, and the base data (VoxPopuli) is already publicly released under CC0. Only the mouth-ROI derivative is distributed — no full faces, no audio, no names/speaker identities. Access is gated to record use and ensure attribution. If you are an individual depicted and wish a clip removed, open an issue.
Limitations
- A handful of clips are very short (< 2 s).
- A/V alignment relies on a per-session offset table; 2020 sittings were accepted at offset 0 (verified by spot-checks) rather than individually audited.
- Italian is 2019-only (VoxPopuli's Italian annotations end in 2019).
- Downloads last month
- 84