README.md

AVSpeech Metadata Files

This repository contains the metadata CSV files for the AVSpeech dataset by Google Research.

Dataset Description

AVSpeech is a large-scale audio-visual speech dataset containing over 290,000 video segments from YouTube, designed for audio-visual speech recognition and lip reading research.

Files

• avspeech_train.csv (128 MB) - Training set with 2,621,845 video segments from 270k videos
• avspeech_test.csv (9 MB) - Test set with video segments from a separate set of 22k videos

CSV Format

Each row contains:

YouTube ID, start_time, end_time, x_coordinate, y_coordinate

Where:

• YouTube ID: The YouTube video identifier
• start_time: Start time of the segment in seconds
• end_time: End time of the segment in seconds
• x_coordinate: X coordinate of the speaker's face center (normalized 0.0-1.0, 0.0 = left)
• y_coordinate: Y coordinate of the speaker's face center (normalized 0.0-1.0, 0.0 = top)

The train and test sets have disjoint speakers.

Usage

With Hugging Face Hub

from huggingface_hub import hf_hub_download

# Download train CSV
train_csv = hf_hub_download(
    repo_id="bbrothers/avspeech-metadata",
    filename="avspeech_train.csv",
    repo_type="dataset"
)

# Download test CSV
test_csv = hf_hub_download(
    repo_id="bbrothers/avspeech-metadata",
    filename="avspeech_test.csv",
    repo_type="dataset"
)

With our dataset loader

from ml.data.av_speech.dataset import AVSpeechDataset

# Initialize dataset (will auto-download CSVs if needed)
dataset = AVSpeechDataset()

# Download videos
dataset.download(
    splits=['train', 'test'],
    max_videos=100,  # Or None for all videos
    num_workers=4
)

Citation

If you use this dataset, please cite the original AVSpeech paper:

@inproceedings{ephrat2018looking,
  title={Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation},
  author={Ephrat, Ariel and Mosseri, Inbar and Lang, Oran and Dekel, Tali and Wilson, Kevin and Hassidim, Avinatan and Freeman, William T and Rubinstein, Michael},
  booktitle={ACM SIGGRAPH 2018},
  year={2018}
}

Links

• AVSpeech Official Page
• Original Paper
• Our GitHub Repository

Notes

• This repository only contains the metadata CSV files, not the actual video content
• Videos must be downloaded from YouTube using the provided YouTube IDs
• Some videos may no longer be available (deleted, private, or geo-blocked)
• Estimated total dataset size: ~4500 hours of video