Initial dataset card

README.md

@@ -0,0 +1,31 @@
+# Dataset: MultiVSR
+
+We introduce a large-scale multilingual lip-reading dataset: MultiVSR. The dataset comprises a total of ~12,000 hours of video footage, covering English + 12 non-English languages. MultiVSR is a massive dataset with a huge diversity in terms of the speakers as well as languages, with approximately 1.6M video clips across 123K YouTube videos. Please check the [https://www.robots.ox.ac.uk/~vgg/research/multivsr/](website) for samples.
+
+<p align="center">
+  <img src="dataset_teaser.gif" alt="MultiVSR Dataset Teaser">
+</p>
+
+
+## Download instructions
+
+Download the list of YouTube IDs, metadata and the train-val-test csvs from `huggingface-datasets` 🤗:
+
+```python
+from datasets import load_dataset
+
+# Login using e.g. `huggingface-cli login` to access this dataset
+ds = load_dataset("sindhuhegde/multivsr")
+```
+
+# Download and Preprocess the videos using the metadata
+
+Follow the instructions in the [https://github.com/Sindhu-Hegde/multivsr/tree/master/dataset](GitHub repo) to download and preprocess the videos.
+
+Once preprocessed, you should have the video clips (`.mp4`) and the transcript (`.txt`) files in the following structure:
+```
+data_root (path of the pre-processed videos) 
+├── list of video-ids
+│ ├── *.mp4 (extracted face track video for each sample)
+|	├── *.txt (full transcript for each clip)
+```