YouTube Subtitles

YT_subtitles is a remarkable tool designed for building a dataset from YouTube subtitles. Let me break it down for you:

  1. Purpose: The primary goal of this tool is to extract non-machine-generated subtitles from YouTube videos. These subtitles are obtained by searching for specific terms and collecting the relevant video content.

  2. How It Works:

    • You provide a list of search terms (such as "movie review," "GPT-3," or "true crime documentary").
    • The tool retrieves videos related to these search terms.
    • For each video, it extracts the subtitles (in various languages) and organizes them into minute-by-minute segments.
    • The resulting files contain a string of text per language, with the language name included as a header.

  3. Dataset Format:

    • The dataset is stored in a JSONL (JSON Lines) file format.
    • Each entry corresponds to a minute of subtitles, with language-specific content.
    • If only one language is available, the output consists of a plain text version of the subtitles without additional metadata.

  4. Use Cases:

    • Researchers and developers can utilize this dataset to enhance the multilingual performance of language models.
    • It's particularly valuable for training models that work with diverse languages and real-world video content.

(1) sdtblck/youtube_subtitle_dataset: YT_subtitles - GitHub. https://github.com/sdtblck/youtube_subtitle_dataset. (2) Youtubean Dataset | Papers With Code. https://paperswithcode.com/dataset/youtubean. (3) youtube subtitles | Kaggle. https://www.kaggle.com/datasets/wadzim/youtube-subtitles. (4) YouTube-8M Segments Dataset - Google Research. https://research.google.com/youtube8m/.