LRS3-TED: a large-scale dataset for visual speech recognition
Paper • 1809.00496 • Published
Datasets:
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 81, in _split_generators
first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 32, in _get_pipeline_from_tar
fs: fsspec.AbstractFileSystem = fsspec.filesystem("memory")
~~~~~~~~~~~~~~~~~^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 302, in filesystem
cls = get_filesystem_class(protocol)
File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 239, in get_filesystem_class
raise ValueError(f"Protocol not known: {protocol}")
ValueError: Protocol not known: memory
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 66, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
LRS3 trainval — auto_avsr mouth crops, bridged to WebDataset
The LRS3 trainval split (30 h) as [T,96,96] uint8 grayscale mouth-ROI shards with character
transcripts, produced with auto_avsr's own crop geometry so results stay comparable to that
project's published LRS3 numbers.
Derived from TheNHz/ellipsis-lrs3-raw
(itself a verified mirror of LRS3-TED, whose official distribution was discontinued).
Attribution
LRS3-TED is by Triantafyllos Afouras, Joon Son Chung and Andrew Zisserman (VGG, University of Oxford), distributed under CC BY 4.0:
T. Afouras, J. S. Chung, A. Zisserman. LRS3-TED: a large-scale dataset for visual speech recognition. arXiv:1809.00496, 2018.
The crop geometry and the published landmarks come from auto_avsr (Pingchuan Ma et al., Imperial College London, Apache 2.0):
P. Ma, A. Haliassos, A. Fernandez-Lopez, H. Chen, S. Petridis, M. Pantic. Auto-AVSR: Audio-visual speech recognition with automatic labels. ICASSP 2023.
Contents
| Clips | 31,957 (of 31,982 — 25 dropped, see below) |
| Duration | 30.07 h @ 25 fps (2,706,466 frames) |
| Clip length | median 2.96 s, min 0.48 s, max 6.20 s |
| Character set | 38 symbols: A–Z, 0–9, space, apostrophe |
| Shards | 32 WebDataset tars, 25.0 GB |
Each sample is three consecutive tar members: <key>.roi.npy ([T,96,96] uint8 grayscale),
<key>.txt (the transcript), <key>.json (provenance: prompt_id, session_key, fps,
n_frames, lang). <key> is a zero-padded index; the real clip identity lives in the JSON.
How the crops were produced
- Source video: LRS3 trainval face tracks (224×224, 25 fps).
- Landmarks: mpc001's published
LRS3_landmarks— no face detection is re-run, so the geometry is auto_avsr's rather than a reimplementation. - Crop: auto_avsr's
VideoProcess(detectors/retinaface/video_process.py) — 68-point similarity warp to their mean face, 96×96 mouth patch,window_margin=12landmark smoothing. - Grayscale: applied after the warp with BT.601 weights (0.299/0.587/0.114), matching
auto_avsr's published order — their preprocessing warps in RGB and applies
torchvision.transforms.Grayscale()at training time. Storing uint8 quantises once (≤1/255).
To feed a model trained on auto_avsr's pipeline, apply their serve transform:
/255 → CenterCrop(88) → Normalize(0.421, 0.165).
25 clips are absent because auto_avsr's VideoProcess returns no patch for them (fewer frames
than the smoothing window, or degenerate landmark geometry) — the same clips its own preprocessing
loop skips. Their IDs are in bridge-report.json.
Verification
- Every clip loads to
[T,1,88,88]float32 through the serve transform. - The auto_avsr
vsr_trlrs3_base.pthvisual encoder yields finite, non-collapsed embeddings: across 100 sampled clips the least-varied has mean pairwise cosine distance 0.745 between frames (a collapsed encoder would sit near 0) — the crops are in-distribution for it. - Shard bytes are reproducible: identical
version_hashacross runs with different worker counts.
dataset-manifest.jsonl carries the content hash
sha256:3534d9b915cff59d1ca38166f19da5d2e7eabb02eea4c0a429d91b9680fd06a3.
Caveats
- This is trainval only — the 407 h pretrain split is not included. Compare against published numbers for the equivalent low-resource (30 h) setting, not against full-corpus results.
- Transcripts are LRS3's own, parsed exactly as auto_avsr parses them (braces stripped, words kept).
- No audio: these shards are for visual-only recognition.
- Downloads last month
- 22