You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Mumospee V2: A MUltiMOdal SPEEch Corpus

Dataset Description

Mumospee v2 is a large-scale multimodal speech dataset containing audio and video recordings with transcripts, aggregated from:

  1. European Council events — official speeches, interviews, and doorstep appearances.
  2. Public YouTube meetings — webinars, conferences, panel discussions, and institutional videos.

This release is built upon the first version of the Mumospee dataset, significantly expanding both the scale and diversity of content, and unifying audio and video data into a consistent metadata format suitable for large-scale speech and multimodal research.

The original Mumospee v1 dataset is available at:
👉 https://huggingface.co/datasets/meetween/mumospee

The dataset is designed for research in speech recognition, multimodal modeling, meeting analysis, and AI-driven content understanding.

All media are referenced via source URLs; no audio or video files are redistributed. Transcripts and metadata are included.

Dataset Structure

Column Type Description
audio_url string URL to audio
video_url string URL to video
duration float Duration in seconds
language string Primary spoken language
transcript string (JSON) Full transcript as JSON-encoded array of segments. Each segment contains language, start_seconds, end_seconds, and text fields with time-aligned transcription data.
split string Source tag (EU_Council or YouTube_Meeting)
license string License information for the content
attribution string (JSON) Attribution as required by the license

Statistic Summary

The dataset is provided as two subsets: eu_council and youtube_meeting.

Source / Tag Number of Records Total Duration (hours) Average Duration (hours)
eu_council 45,411 3,665.68 0.0807
youTube_meeting 122,407 80,127.65 0.6546

Language Distribution

The table below reports the primary spoken language per recording.

Language EU_Council Count EU_Council Duration (hours) YouTube_Meeting Count YouTube_Meeting Duration (hours)
English 23635 1754.21 95896 59568.38
German 3582 294.16 1467 831.99
French 2953 240.68 8222 5910.38
Spanish 2427 230.31 11382 9497.5
Italian 1374 122.0 1471 1092.87
Portuguese 1230 113.16 3030 2706.79
Swedish 1092 96.51 82 49.71
Dutch 1064 97.09 782 418.2
Polish 970 91.22 29 24.79
Czech 873 77.6 13 8.19
Croatian 826 71.11 2 2.96
Danish 787 71.06 5 0.95
Slovak 684 62.17 1 1.15
Finnish 647 62.9 5 5.36
Greek 604 37.16 13 4.87
Slovenian 519 40.49 0 0.0
Bulgarian 401 66.11 2 0.62
Hungarian 384 34.39 2 0.48
Luxembourgish 297 28.92 0 0.0
Romanian 295 25.66 1 0.07
Maltese 156 5.63 0 0.0
Lithuanian 92 5.32 0 0.0
Arabic 79 4.53 0 0.0
Latvian 73 5.18 0 0.0
Ukrainian 56 4.48 0 0.0
Russian 38 2.67 0 0.0
Estonian 37 2.36 0 0.0
Serbian 36 2.84 0 0.0
Georgian 33 1.99 0 0.0
Norwegian 29 2.98 0 0.0
Albanian 25 1.89 0 0.0
Macedonian 23 1.78 0 0.0
Bosnian 17 1.36 0 0.0
Montenegrin 16 1.09 0 0.0
Turkish 14 1.37 0 0.0
Belarusian 6 0.31 0 0.0
Moldavian 5 0.34 0 0.0
Catalan 4 0.41 0 0.0
Persian 4 0.36 0 0.0
Japanese 4 0.25 0 0.0
Indonesian 3 0.25 0 0.0
Vietnamese 3 0.33 0 0.0
Chinese 3 0.28 0 0.0
Armenian 3 0.07 0 0.0
Korean 3 0.33 0 0.0
Hindi 1 0.1 0 0.0
Kazakh 1 0.1 0 0.0
Khmer 1 0.06 0 0.0
Tajik 1 0.09 0 0.0
Swahili 1 0.05 0 0.0

Notes on Transcripts

  • EU Council (audio & video): Transcripts are generated using the Whisper ASR package and are not official EU transcripts. Each entry contains one or more languages per recording, corresponding to the full speech.

  • YouTube Meeting (video):
    Transcripts are extracted from YouTube subtitle tracks. They may be auto-generated or user-uploaded, marked in the metadata. Transcript quality may therefore vary.


Usage

from datasets import load_dataset

# Load the dataset in streaming mode (recommended for large datasets)
dataset = load_dataset("meetween/mumospee_v2", streaming=True)


# Access first row of from eu_council set
first_row = next(iter(dataset["eu_council"]))
print("Transcript:", first_row["transcript"])
print("Audio URL:", first_row["audio_url"])
print("Video URL:", first_row["video_url"])
print("Language:", first_row["language"])
print("Duration (s):", first_row["duration"])


# Iterate over the first 5 examples
for i, example in enumerate(dataset["eu_council"]):
    print(f"{i+1}. {example['language']} - {example['duration']}s")
    print(f"Transcript snippet: {example['transcript'][:100]}...\n")
    if i >= 4:
        break

# Access audio or video URLs
first_example = next(iter(dataset["eu_council"]))
print("Listen to audio:", first_example["audio_url"])
print("Watch video:", first_example["video_url"])

Licensing Information

YouTube CC-BY:
All YouTube content in Mumospee v2 is selected from videos marked with the Creative Commons Attribution (CC BY) license. CC BY permits reuse, transformation, and training uses as long as proper attribution is maintained (e.g., title, uploader, source URL). Users must comply with the CC BY terms when accessing media via URLs. Creative Commons

European Council reuse notice:
Council of the EU content is published with a copyright notice stating that reproduction is authorised provided the original source is acknowledged and the meaning is not distorted. Mumospee distributes derived metadata and signal information only; raw media remains at the source and is accessed by URL. European Council

Users must comply with the source license or notice when accessing or reusing media via URLs.

Intended Use and Data Quality

Mumospee v2 is intended for speech and multimodal research, including speech recognition, meeting analysis, and language understanding. The current dataset includes metadata and filtered based on LLM prompting. As such, some videos may be incorrectly labeled for meetings or have misidentified languages. In future releases we will incorporate audio and video analysis pipelines to refine classification and release timestamped segments for downstream tasks.

Downloads last month
119