meetween/mumospee_v3.1-testing

Mumospee V3

Dataset summary

Mumospee V3 is a speech corpus of 6,448,108 sentence-level clips cut from 46,373 YouTube videos in 43 languages, 29,751 hours in total. The videos come from the YouTube collection of Mumospee V2 (youtube_meeting split): 33,275 of the 46,373 videos (71.8%) are in that split, and the others are videos of the same collection that are not included in it.

The corpus is organised in subsets, one per task. Every subset has the columns that describe the clip, and in addition only the columns of its task. Every row of a clip-level subset refers to a clip of the corpus through clip_id, so the rows of different tasks on the same clip can be joined. The audio is not included in this repository; audio_input_path gives the relative path of each clip's audio file (see Audio files).

All source videos were published under the Creative Commons Attribution license. The license of every video was checked with the YouTube Data API on 2026-10-15; videos that were no longer published under this license on that date are not included.

This is release v1: it contains the videos in languages other than English and a sample of about 8,000 hours of English videos. The remaining English videos are planned for a later release.

Subsets

Subset	Task	Rows	Videos	Hours	Languages	Columns	Files
asr	Automatic speech recognition (ASR)	6,448,108	46,373	29,751.0	43	22	data/asr/<lang>/

Each subset is one configuration with all its languages in the split train; the first subset is the default configuration. The files of each language are in their own folder (data/<task>/<lang>/), so one language can be loaded on its own by selecting its files (see Usage).

Usage

from datasets import load_dataset

# one subset, all languages, streamed
ds = load_dataset("meetween/mumospee_v3.1-testing", "asr", split="train", streaming=True)

# one subset, one language: select the files of the language
ds = load_dataset("meetween/mumospee_v3.1-testing", data_files="data/asr/nl/*.parquet", split="train")
row = next(iter(ds))
print(row["clip_id"], row["video_id"], row["start"], row["end"], row["text_output"])

The source video of a clip is https://www.youtube.com/watch?v=<video_id>; the clip is the interval from start to end seconds.

Dataset structure

data/<task>/<lang>/train-NNNNN-of-MMMMM.parquet    one folder per subset and language
scripts/make_clips.py                              cuts the clip audio from source videos

Audio files

The audio files are not part of this repository. audio_input_path is the relative path of the clip audio (16 kHz mono FLAC), the same in every clip-level subset:

audio/<lang>/<video_id>/<clip_id>.flac      e.g. audio/nl/-H9G_NHn4Xc/-H9G_NHn4Xc_0000.flac

<lang> is the video language. If the folder that contains audio/ is <root>, the audio of a row is <root>/<audio_input_path>. The script scripts/make_clips.py cuts the clips from source videos that you have downloaded and writes them at these paths; see scripts/README.md.

Columns of every subset

These columns describe the clip and have the same values in every subset that contains the clip.

Column	Type	Value in this dataset	Explanation
clip_id	string	<video_id>_<index>	ID of the clip; the index is the position of the clip in its video, from 0. The rows of all subsets that are built on the same clip have the same clip_id
video_id	string	YouTube video ID	Source video: https://www.youtube.com/watch?v=<video_id>
start	float64	seconds	Start of the clip in the source video
end	float64	seconds	End of the clip in the source video
audio_input_path	string	audio/<lang>/<video_id>/<clip_id>.flac	Relative path of the clip audio (16 kHz mono FLAC); the audio files are not part of this repository (see Audio files)
audio_input_duration	float64	seconds, e.g. 15.75	Duration of the clip
lid_audio_predicted	string	ISO 639-1 code or null	Spoken language detected by Whisper large-v3-turbo from the audio of the source video (first 30 seconds of speech); one value per video, not per clip. Null for the videos of the first processing batch, which were transcribed with the language set in advance, so no detection was made (2,819,246 rows, 43.7%)
language_confidence_audio	float64	0 to 1 or null	Probability of lid_audio_predicted given by Whisper
license	string	CC-BY-4.0	License of the row
attribution	string	title, uploader, channel_url, video_id	JSON string with the title, channel name, channel URL and ID of the source video, for the attribution required by the CC BY license

Column names and values that are not stored

The column names and types are those of the dataset format of speechLMM-v2, a shared format for speech and multimodal training data. Each subset contains only the columns that its task uses. The following values are the same for every row and are therefore not stored as columns: the clip audio is 16 kHz, mono, FLAC; the transcripts are model output, not human references; all rows are training data (split train).

speechLMM-v2 users do not need to convert the data themselves: its mumospee_v3 corpus module reads a subset, adds the other columns of the speechLMM format, and extracts the clip audio.

Subset asr: Automatic speech recognition (ASR)

One row per clip: the clip audio is the input, the Whisper large-v3-turbo transcript of the clip (cased and punctuated) is the target (text_output).

Files: data/asr/<lang>/train-NNNNN-of-MMMMM.parquet, at most 500,000 rows per file, 51 files. The language of a folder is the language of the video, determined from the language labels of its clips; individual clips can be in another language (code-switching), given per clip in lid_text_predicted.

Language	Files	Rows	Videos	Hours
Spanish (es)	data/asr/es/	1,821,626	11,319	8,675.6
English (en)	data/asr/en/	1,902,083	16,778	8,537.1
French (fr)	data/asr/fr/	1,091,144	7,465	5,019.6
Russian (ru)	data/asr/ru/	449,162	2,993	2,027.0
Portuguese (pt)	data/asr/pt/	431,715	2,411	2,021.8
German (de)	data/asr/de/	259,535	1,921	1,193.7
Italian (it)	data/asr/it/	194,942	1,332	972.9
Indonesian (id)	data/asr/id/	121,149	619	531.6
Dutch (nl)	data/asr/nl/	76,344	609	323.4
Ukrainian (uk)	data/asr/uk/	17,296	96	82.1
Swedish (sv)	data/asr/sv/	15,552	113	67.6
Korean (ko)	data/asr/ko/	12,791	139	60.1
Turkish (tr)	data/asr/tr/	11,933	82	52.9
Hindi (hi)	data/asr/hi/	7,535	168	30.8
Vietnamese (vi)	data/asr/vi/	5,643	69	24.4
Norwegian (no)	data/asr/no/	3,918	23	18.0
Arabic (ar)	data/asr/ar/	4,192	26	16.9
Catalan (ca)	data/asr/ca/	2,936	24	13.9
Polish (pl)	data/asr/pl/	2,350	23	11.2
Hebrew (he)	data/asr/he/	2,299	31	10.7
Greek (el)	data/asr/el/	2,859	15	9.9
Japanese (ja)	data/asr/ja/	1,978	31	8.9
Kazakh (kk)	data/asr/kk/	1,753	9	7.4
Czech (cs)	data/asr/cs/	1,585	13	7.3
Finnish (fi)	data/asr/fi/	1,497	8	7.0
Tagalog (tl)	data/asr/tl/	1,618	12	6.7
Galician (gl)	data/asr/gl/	610	6	2.9
Serbian (sr)	data/asr/sr/	533	3	2.5
Malay (ms)	data/asr/ms/	290	2	1.4
Bulgarian (bg)	data/asr/bg/	254	7	1.2
Azerbaijani (az)	data/asr/az/	200	4	0.9
Danish (da)	data/asr/da/	170	5	0.8
Hungarian (hu)	data/asr/hu/	106	2	0.5
Slovak (sk)	data/asr/sk/	143	1	0.5
Slovenian (sl)	data/asr/sl/	73	1	0.3
Tamil (ta)	data/asr/ta/	70	2	0.3
Swahili (sw)	data/asr/sw/	62	3	0.3
Urdu (ur)	data/asr/ur/	53	2	0.2
Welsh (cy)	data/asr/cy/	46	2	0.2
Estonian (et)	data/asr/et/	25	1	0.1
Romanian (ro)	data/asr/ro/	15	1	0.1
Gujarati (gu)	data/asr/gu/	11	1	0.1
Marathi (mr)	data/asr/mr/	12	1	0.1
all languages	data/asr/*/	6,448,108	46,373	29,751.0

Columns of this subset, in addition to the columns of every subset:

Column	Type	Value in this dataset	Explanation
text_output	string	transcript	Target text: the Whisper large-v3-turbo transcript of the clip, cased and punctuated
text_output_normalized	string	lower-cased transcript	text_output in lower case, with whitespace collapsed to single spaces; punctuation is kept
num_words_output	int64	e.g. 37	Number of words in text_output
language_text_output	string	ISO 639-1 code or null	Language of text_output: the label of lid_text_predicted, or null if no language was identified (28,212 rows, 0.4%), so that tools treat the language as unknown instead of as a language code. It is also the language of the input audio
lid_text_predicted	string	ISO 639-1 code or und	Output of the language classifier (fastText lid.176) on the transcript. It records the classifier result and is never null: und (ISO 639-2 code for "undetermined") means that the classifier ran but identified no language. These rows also have the flag lang_und
language_confidence_text	float64	0 to 1	Confidence of lid_text_predicted
words_per_second	float64	e.g. 2.35	Speaking rate: num_words_output divided by audio_input_duration
chars_per_second	float64	e.g. 13.8	Characters of text_output per second of audio
repetition_score	float64	0 to 1	Share of repeated word 3-grams in text_output_normalized: 1 − distinct 3-grams / all 3-grams (character 3-grams for Japanese and other languages written without spaces; 0 for texts of 3 tokens or fewer). High values indicate a repetition loop of the speech recognition model; 13,370 rows have a score above 0.5
text_duplicate_score	float64	0.0 or 1.0	1.0 if another row with a smaller clip_id has the same text_output_normalized (123,799 rows), otherwise 0.0
text_duplicate_of	string	clip_id or null	For rows with text_duplicate_score 1.0, the clip_id of the first row (smallest clip_id) with the same text; null otherwise
flags	list of strings	empty, fallback_cut, lang_und	Quality warnings. fallback_cut: no sentence ended within 30 seconds, so the clip ends at the longest pause instead of a sentence end. lang_und: the language of the transcript could not be identified

Example row. The transcript is shortened here, and attribution, which is stored as a JSON string, is shown parsed.

{
  "clip_id": "-H9G_NHn4Xc_0000",
  "video_id": "-H9G_NHn4Xc",
  "start": 0.0,
  "end": 15.75,
  "audio_input_path": "audio/nl/-H9G_NHn4Xc/-H9G_NHn4Xc_0000.flac",
  "audio_input_duration": 15.75,
  "text_output": "In deze video laat ik zien hoe je GA4 e-commerce setting instelt aan de hand van Tag Manager. Je moet helaas zelf gebeurtenissen toevoegen. In dit voorbeeld mak ...",
  "text_output_normalized": "in deze video laat ik zien hoe je ga4 e-commerce setting instelt aan de hand van tag manager. je moet helaas zelf gebeurtenissen toevoegen. in dit voorbeeld mak ...",
  "num_words_output": 37,
  "language_text_output": "nl",
  "lid_text_predicted": "nl",
  "language_confidence_text": 0.988,
  "words_per_second": 2.3492063492063493,
  "chars_per_second": 13.841269841269842,
  "repetition_score": 0.0,
  "text_duplicate_score": 0.0,
  "text_duplicate_of": null,
  "flags": [],
  "lid_audio_predicted": "nl",
  "language_confidence_audio": 0.999,
  "license": "CC-BY-4.0",
  "attribution": {
    "title": "E-commerce instellen in Google Analytics 4",
    "uploader": "Altijd Bekend",
    "channel_url": "https://www.youtube.com/channel/UC0ILpCCSKhitmVEp9ykZKsg",
    "video_id": "-H9G_NHn4Xc"
  }
}

Licensing and attribution

The dataset is released under the Creative Commons Attribution 4.0 license (CC-BY-4.0). Each source video was published on YouTube under the Creative Commons Attribution license on 2026-10-15. When you use or redistribute the data, credit the authors of the source videos; the attribution column gives the title, channel name and channel URL of each video.

Uploaders can change the license of a video or remove it after the check. The license status refers to the date of the check.