TheNHz/ellipsis-lrs3-raw

LRS3-TED — verified mirror

A mirror of the LRS3-TED dataset (Lip Reading Sentences 3), preserved because the official distribution has been discontinued. This repository adds no new data: it is a re-hosted copy with a full verification report against the official file list, so you know exactly what is and is not here.

Attribution

LRS3-TED was created by Triantafyllos Afouras, Joon Son Chung and Andrew Zisserman (Visual Geometry Group, University of Oxford):

T. Afouras, J. S. Chung, A. Zisserman. LRS3-TED: a large-scale dataset for visual speech recognition. arXiv:1809.00496, 2018.

The dataset is distributed under CC BY 4.0 (as stated on the official dataset page, mmai.io/datasets/lip_reading/). The underlying material consists of TED and TEDx talks; please also respect TED's own terms for the source media.

Status of the official distribution

As of 2026-07, the official downloads are gone: mmai.io/datasets/lip_reading/ states "Downloads are no longer available from this website", the Oxford VGG page returns 404, and the thor.robots.ox.ac.uk archives return 404. One of the authors (Prof. Zisserman) has confirmed by email to other researchers that LRS3 is no longer available. This mirror exists so that published LRS3 results remain reproducible.

Contents

Path	Contents
ainncy/pretrain.00–99.tar	pretrain split, original layout pretrain/{youtube_id}/{clip}.mp4 + .txt
ainncy/trainval.tar.gz	trainval split, same layout
test-mattymchen/	test split (1,321 utterances) as parquet: audio, video, transcript
landmarks/LRS3_landmarks.zip	precomputed facial landmarks (auto_avsr/mpc001 preprocessing)
verification/	verification reports, file listings, missing-clip lists

Media format (verified): 224×224 face-track video, MPEG-4 Part 2, 25 fps, with AAC 16 kHz audio. Transcripts are the original LRS3 .txt format (Text: / Conf: / per-word WORD START END ASDSCORE).

Verification — read this before using

Counted against the official file list (recovered from LRS3_landmarks.zip, which enumerates all 151,819 clips):

Split	Clips here	Official	Status
pretrain	103,491 unique	118,516	87.3 % — 636 whole videos absent
trainval	31,982	31,982	complete
test	1,321	1,321	complete

Independent identity check: all 1,200/1,200 validation IDs from Meta's av_hubert (lrs3-valid.id) are present in trainval. Authenticity spot-check: sampled YouTube IDs resolve to TED/TEDx channel videos whose content matches the accompanying transcripts.

Two caveats:

1. Missing pretrain videos. 636 videos are absent entirely (never partially), in contiguous alphabetical bands of the YouTube-ID ordering — the loss pattern of a repack, not of corruption. Exact lists: verification/missing_pretrain_{clips,videos}.lst.
2. Duplicate clips at tar boundaries. The pretrain tars contain 2,360 clips twice, so a naive listing counts 105,851 entries for 103,491 unique clips. Deduplicate on (youtube_id, clip_id) when building an index, or you will train on duplicated samples.

The trainval and test splits are complete, so benchmark numbers on them remain comparable to published results. Pretrain-split results are not strictly comparable to work using the full corpus.

Provenance

Re-staged from the public mirror Ainncy/LRS3 (pretrain + trainval), mattymchen/lrs3-test (test split) and faori/lrs3_landmark (landmarks), then verified as described above. Verification tooling and reports are included in verification/.