Remove dataset card
Browse files
README.md
DELETED
|
@@ -1,71 +0,0 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: other
|
| 3 |
-
license_name: lrs3-ted-derived
|
| 4 |
-
license_link: https://mmai.io/datasets/lip_reading/
|
| 5 |
-
pretty_name: LRS3 preprocessed features (LipCFM)
|
| 6 |
-
task_categories:
|
| 7 |
-
- audio-to-audio
|
| 8 |
-
- automatic-speech-recognition
|
| 9 |
-
language:
|
| 10 |
-
- en
|
| 11 |
-
size_categories:
|
| 12 |
-
- 100B<n<1T
|
| 13 |
-
---
|
| 14 |
-
|
| 15 |
-
# LRS3 preprocessed — audio, mels, mouth ROIs, face crops, AV-HuBERT & ArcFace features
|
| 16 |
-
|
| 17 |
-
Preprocessed derivatives of **LRS3-TED** used to train LipCFM (lip-conditioned conditional
|
| 18 |
-
flow matching, F5-TTS style). 302,167 utterances / 438 h.
|
| 19 |
-
|
| 20 |
-
> **Provenance / rights.** These are derivatives of LRS3-TED (Oxford VGG). The uploader
|
| 21 |
-
> states that redistribution permission was obtained from the dataset authors. Downstream
|
| 22 |
-
> users remain bound by the original LRS3-TED terms — check before reuse.
|
| 23 |
-
|
| 24 |
-
## Contents
|
| 25 |
-
|
| 26 |
-
Total 322.3 GB in 30 parts. Excluded from this archive: `landmark/` (21.5 GB) and the
|
| 27 |
-
`records*/` provenance directories.
|
| 28 |
-
|
| 29 |
-
| Directory | Size | Format |
|
| 30 |
-
|---|---|---|
|
| 31 |
-
| `audio/` | 50.63 GB | 16 kHz mono wav |
|
| 32 |
-
| `audio_norm/` | 50.63 GB | same, loudness-normalised to −23 LUFS |
|
| 33 |
-
| `mel/` | 50.72 GB | `(T,80)` f32 Tacotron2 log-mel, hop 160 / n_fft 640 / fmax 8000 |
|
| 34 |
-
| `mel_norm/` | 50.72 GB | log-mel of `audio_norm/` — the CFM training target |
|
| 35 |
-
| `video/` | 12.64 GB | 96×96 colour mouth ROI, x264 crf 20 @25 fps |
|
| 36 |
-
| `face/` | 58.67 GB | 224×224 aligned face crop, x264 crf 18 @25 fps |
|
| 37 |
-
| `avhubert_slv433_L24/` | 80.79 GB | `feat.f16.npy (ΣT,1024)` f16 + offsets/nframes/ids/meta/stats |
|
| 38 |
-
| `spk/` | 1.26 GB | `emb.f32.npy (N,512)` ArcFace utterance embeddings |
|
| 39 |
-
| `emo_poster_*/` | 6.13 GB | POSTER++ per-frame emotion logits + featmean |
|
| 40 |
-
| `label*/` | 0.11 GB | fairseq tsv manifests + `.wrd` transcripts |
|
| 41 |
-
|
| 42 |
-
Alignment: video is 25 fps, mel 100 Hz — 4 mel frames per video frame. ~50% of train
|
| 43 |
-
utterances carry a 0–40 ms audio/video start offset; see the training repo's
|
| 44 |
-
`lipcfm/data/manifest.py` for the correction.
|
| 45 |
-
|
| 46 |
-
## Layout of the parts
|
| 47 |
-
|
| 48 |
-
- `lrs3.files.part00..22.tar.gz` — 23 independent tarballs, 1,813,241 files
|
| 49 |
-
- `lrs3.feat.part00..06.gz` — 7 raw byte chunks of the single 80.76 GB
|
| 50 |
-
`avhubert_slv433_L24/feat.f16.npy`, which cannot be split across tar members
|
| 51 |
-
|
| 52 |
-
## Restore
|
| 53 |
-
|
| 54 |
-
```bash
|
| 55 |
-
huggingface-cli download <REPO_ID> --repo-type dataset --local-dir lrs3_parts
|
| 56 |
-
cd lrs3_parts && ./restore.sh /destination
|
| 57 |
-
```
|
| 58 |
-
|
| 59 |
-
`restore.sh` verifies all 30 checksums, extracts the tarballs, reassembles the feature
|
| 60 |
-
file, then asserts 1,813,241 files and `feat.f16.npy` == 80,764,827,776 bytes.
|
| 61 |
-
|
| 62 |
-
Manual equivalent:
|
| 63 |
-
|
| 64 |
-
```bash
|
| 65 |
-
for f in lrs3.files.part*.tar.gz; do tar -xzf "$f" -C /destination; done
|
| 66 |
-
cat lrs3.feat.part*.gz | gunzip > /destination/lrs3/avhubert_slv433_L24/feat.f16.npy
|
| 67 |
-
```
|
| 68 |
-
|
| 69 |
-
## Training code
|
| 70 |
-
|
| 71 |
-
https://github.com/longkhanh-fam/interspeech2027
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|