williamTLmiller/youtube-transcripts-creative-commons

YouTube Video Transcripts Dataset: Creative Commons, Multilingual (CSV Sample)

Release name: YouTube-Commons Normalized v1: free CSV sample

A free, CSV-only sample of a normalized, de-duplicated and labeled build of YouTube-Commons: speech transcripts of YouTube videos that their creators published under the CC BY license.

The sample covers 5,953 channels (about 1% of all channels, at most 25 videos each), 18,081 videos and 21,010 transcripts. For each sample video it includes the primary transcript (English when available) and the original-language transcript when that differs.

Full and trainable versions

Package	Contents	Format
Free sample (this repo)	~1% of channels	CSV
Trainable	3,112,197 videos with a non-empty transcript, one primary transcript each, train/validation/test 98/1/1	Parquet
Full	3,156,666 videos, 585,789 channels, 21,736,981 transcripts in 128 languages	Parquet + CSV tables

The trainable and full versions are commercial products. Contact energea@energeatech.com.

Files

• csv/videos.csv, csv/channels.csv: tables described below
• csv/transcripts.csv: the transcripts, with channel and video_url repeated on each row for attribution

videos (one row per video)

video_id, video_url, title, channel, channel_id, channel_url, upload_date, upload_year, source_license, original_language, language_id_method, primary_transcript_language, primary_is_machine_translation, primary_word_count, topic, topic_confidence, format, quality_flags, transcript_languages, transcript_count, has_original_language_transcript, split

channels (one row per channel)

channel_id, channel, channel_url, video_count, first_upload_date, last_upload_date, main_original_language, main_topic, total_primary_words, split

transcripts (one row per video and language)

video_id, channel, video_url, language, is_original_language, is_machine_translation, word_count, character_count, quality_flags, split, text

Labels (silver)

Deterministic, rule-derived labels, reproducible from the public fields. They are not human-verified.

• topic (from the title plus the English transcript): education, science_technology, news_politics, government_public_meetings, religion_spirituality, health_medicine, business_finance, gaming, music, sports_fitness, food_cooking, travel_outdoors, autos_vehicles, how_to_diy, arts_culture_history, kids_family, other, unknown. unknown means no English evidence; other means weak evidence. topic_confidence is the winning topic's share of evidence.
• format (from title and channel name): tts_narration, public_meeting, sermon_worship, lecture_talk, interview_podcast, news_report, tutorial_howto, review_unboxing, gameplay, music_performance, livestream, vlog, other.
• quality_flags (semicolon-separated, empty when clean): empty_transcript, very_short, mostly_music_or_noise, highly_repetitive, language_code_repaired.
• split: train / validation / test = 98 / 1 / 1, assigned by channel, so no creator appears in more than one split.

License

• The MatrixNorm packaging layer (tables, labels, splits, documentation) is licensed CC BY-NC 4.0 (williamTLmiller), see LICENSE.
• The transcripts and video metadata themselves remain CC BY 4.0 from the original creators via PleIAs YouTube-Commons, see NOTICE and LICENSE-CC-BY-4.0-UNDERLYING-DATA.txt. Keep the creator attribution columns when you reuse them.

Source and limitations

• Source: PleIAs/YouTube-Commons (CC BY 4.0).
• Transcripts are automatic speech recognition and automatic translation output and contain errors.
• E-mail addresses and phone numbers in transcripts are masked as [EMAIL] / [PHONE].
• License status is as declared by each uploader on YouTube.
• A small number of source videos (under 0.1%) have no channel name or title in the source; their video URL is kept as the attribution link.
• Labels are silver: use them as weak supervision, not ground truth.

Contact: energea@energeatech.com