iDRAMA-rumble-2024: A Dataset of Podcasts from Rumble Spanning 2020 to 2022

ABSTRACT

---------------
Rumble has emerged as a prominent platform hosting controversial figures facing restrictions on YouTube. Despite this, the academic community’s engagement with Rumble has been minimal. To help researchers address this gap, we introduce a comprehensive dataset of about 6.7K podcast videos from August 2020 to December 2022, amounting to over 5.6K hours of content. Besides covering metadata of these podcast videos, we provide speech-to-text transcriptions for future analysis. We also provide speaker diarization information, a collection of ~250K unique representative images from podcast videos, and face embeddings of ~400K extracted faces. With the rise of the influence of podcasts and populist figures, this dataset provides a rich resource for identifying challenges in cyber social threats in a relatively underexplored space.

• Rumble platform: http://rumble.com/
• Link to paper: https://workshop-proceedings.icwsm.org/abstract.php?id=2024_07
• License: CC BY-NC-SA 4.0

Dataset Summary

iDRAMA-rumble-2024 is a large-scale dataset of 6,735 podcast videos from Rumble, an alternative Youtube-like platform. Using state-of-the-art models, we extract information across three modalities: 1) text, 2) audio, and 3) video. We detail the methodology for extracting information from podcast videos in the paper and release a first-of-its-kind dataset including data from different modalities:

• Metadata: Details about podcast videos, e.g., channel name, video name, video description, and more.
• Text: Transcription (i.e., speech-to-text) of podcast videos.
• Audio: Speaker diarization information providing speaker detection over time for each video.
• Video: Sampled representative video frames from each video, totaling 200K images. We also detect ~400K non-unique faces from these images and release face embeddings.

Repository links

• Zenodo: On Zenodo, we provide JSON formatted dataset for all modalities and representative images in compressed files.
• Github: The main repository of this dataset, where we provide code snippets to get started with this dataset.
  • Link here: https://github.com/idramalab/iDRAMA-rumble-2024
• Huggingface: On Huggingface, we provide a dataset that can be accessed through Huggingface APIs in a `parquet` format.
  • Link here: https://hf.co/datasets/iDRAMALab/iDRAMA-rumble-2024

Dataset Info

The dataset is organized by modalities -- transcripts, representative images, speaker diarization, and face embeddings.

Zenodo Dataset Files Info

Authorship

This dataset is published in the "Workshop Proceedings of the 18th International AAAI Conference on Web and Social Media" hosted in Buffalo, NY, USA.

• Academic Organization: iDRAMA Lab
• Authors: Utkucan Balci, Jay Patel, Berkan Balci, Jeremy Blackburn
• Affiliation: Binghamton University, Middle East Technical University

Licensing

This dataset is available for free to use under terms of the non-commercial license CC BY-NC-SA 4.0.

Citation

@article{balci2024idrama,
  title  = {iDRAMA-rumble-2024: A Dataset of Podcasts from Rumble Spanning 2020 to 2022},
  author = {Balci, Utkucan and Patel, Jay and Balci, Berkan and Blackburn, Jeremy},
  year   = {2024},
  journal = {Workshop Proceedings of the 18th International AAAI Conference on Web and Social Media}
}