Meetween YouTube Meeting Video Dataset
Dataset Description
The Meetween YouTube Meeting Dataset is a comprehensive corpus comprising processed transcripts and metadata from 122,407 public YouTube videos, totaling over 80,000 hours of audiovisual content. Part of the Mumospee v2 release, this dataset is specifically designed to support AI research in meeting analysis, speech processing, and multimodal technologies.
The content focuses on virtual meetings, webinars, conferences, panel discussions, speeches, and institutional communications. Each entry pairs a video with extensive metadata, including automatically extracted transcripts (aggregated from manual and auto-generated subtitles), language identification, and AI-generated classification labels for relevance, speaker count, and topic.
Dataset Structure |
| Column Name | Data Type | Description |
|---|---|---|
video_id |
string | Unique YouTube video identifier |
video_link |
string | Direct link to the YouTube video |
title |
string | Video title |
channel_name |
string | YouTube channel name |
duration_seconds |
int | Duration in seconds |
text |
string | Full transcript text |
license |
string | Video license |
original_language |
string | Primary detected or reported language of the video |
transcripts_languages |
list of string | All languages for which transcript exist |
autogenerated_transcripts |
boolean | Whether transcripts are auto-generated by YouTube |
word_count |
int | Number of words in the transcript |
character_count |
int | Number of characters in the transcript |
description |
string | Full YouTube description |
search_keywords |
string | Query used to retrieve or identify this video |
classification |
struct | Meetween-specific relevance classification metadata |
└─ meetween_relevance_score |
float / int | Model-assigned relevance score |
Dataset Size
- Number of records: 122407
- Total duration in hours: 80122.65
- Average transcript length: 5999.1
Language Distribution
The dataset covers 18 of the 24 official EU languages.
- Dominant Language: English (approx. 78% of the dataset).
- Multilingual Support: Includes a "long tail" of low-resource EU languages.
| Language | Number of videos | Total Duration (hours) |
|---|---|---|
| English | 95,896 | 59,568 |
| Spanish | 11,383 | 9,499 |
| French | 8,223 | 5,910 |
| Portuguese | 3,030 | 2,706 |
| Italian | 1,471 | 1,092 |
| German | 1,467 | 831 |
| Dutch | 782 | 418 |
| Swedish | 82 | 49 |
| Polish | 29 | 25 |
| Czech | 13 | 8.19 |
| Greek | 13 | 4.87 |
| Finnish | 5 | 5.36 |
| Danish | 5 | 0.95 |
| Croatian | 2 | 2.96 |
| Bulgarian | 2 | 0.62 |
| Hungarian | 2 | 0.48 |
| Slovak | 1 | 1.15 |
| Romanian | 1 | 0.07 |
Usage
from datasets import load_dataset
dataset = load_dataset("username/youtube-meeting-videos")
print(dataset['train'][0]['text'])
print(dataset['train'][0]['video_id'])
print(dataset['train'][0]['classification'])
Notes on Transcripts
- Transcripts are extracted from YouTube subtitle tracks.
- Auto-generated subtitles may contain errors.
- Some videos may lack transcripts entirely.
Licensing information
Metadata & Transcripts: Released under the CC0 (Public Domain) license.
Video Content: The original video content is released under the CC BY (Attribution) license by the respective content creators. Users must attribute the original uploader when using the raw audio/visual data.
Note: While this Hugging Face dataset provides the metadata and transcripts, public users should use the video_id to access the source content directly from YouTube if raw media is required.
- Downloads last month
- 19