Mumospee: A MUltiMOdal SPEEch Corpus

The Mumospee dataset supports the Meetween project's mission of enabling inclusive, language-neutral collaboration across virtual environments. The release provides metadata and download URLs for a curated collection of speech audio sourced from publicly available datasets, optimized for processing on high-performance computing clusters.

Mumospee Overview

Mumospee is a comprehensive multilingual speech-metadata corpus featuring:

• 140,680 hours of speech metadata across 54,389,354 samples
• Coverage of 25 EU languages plus a long tail of additional languages
• Collections drawn from existing datasets in different speaking styles and content genres:

_TAGS = ["CoVoST", "GigaSpeech", "PeopleSpeech", "Librispeech", "LibriTTS", "Emilia", "MOSEL"]

A smaller version with fewer than 1000 rows is also available as mumospee_small for testing purposes.

Dataset Statistics

Overview (All Splits Combined)

Metric	Value
Total samples	54,389,354
Total audio duration	140,680h 05m 02.7s (140,680.1h)
Average duration per sample	9.31s
Avg transcript length	16.4 words
Total parquet shards	30

Per-Split Overview

Split	# Samples	Duration	Avg Duration	Avg Words	Shards
train	53,652,580	139,490h 52m 18.3s (139,490.9h)	9.36s	16.5	28
test	378,260	603h 44m 33.9s (603.7h)	5.75s	10.0	1
validation	358,514	585h 28m 10.4s (585.5h)	5.88s	10.2	1

Language Distribution

Value	train samples	train %	test samples	test %	validation samples	validation %	Total samples	Total %	Total Duration	Total Dur %
en	28,924,123	53.91%	282,248	74.62%	263,734	73.56%	29,470,105	54.18%	67,654h 01m 01.6s	48.09%
zh	19,969,304	37.22%	0	0.00%	0	0.00%	19,969,304	36.72%	49,922h 33m 08.9s	35.49%
ja	870,783	1.62%	684	0.18%	635	0.18%	872,102	1.60%	1,718h 30m 51.2s	1.22%
de	824,954	1.54%	13,511	3.57%	13,511	3.77%	851,976	1.57%	2,551h 56m 23.0s	1.81%
fr	774,324	1.44%	14,760	3.90%	14,760	4.12%	803,844	1.48%	2,400h 42m 27.9s	1.71%
es	164,355	0.31%	13,221	3.50%	13,221	3.69%	190,797	0.35%	933h 24m 27.1s	0.66%
it	119,389	0.22%	8,183	2.16%	8,940	2.49%	136,512	0.25%	852h 27m 43.2s	0.61%
pt	109,160	0.20%	4,023	1.06%	3,318	0.93%	116,501	0.21%	805h 56m 44.0s	0.57%
nl	107,110	0.20%	1,699	0.45%	1,699	0.47%	110,508	0.20%	803h 57m 04.0s	0.57%
lv	102,338	0.19%	1,629	0.43%	1,125	0.31%	105,092	0.19%	784h 46m 37.6s	0.56%
et	101,752	0.19%	1,571	0.42%	1,576	0.44%	104,899	0.19%	791h 53m 08.0s	0.56%
sl	101,568	0.19%	360	0.10%	509	0.14%	102,437	0.19%	786h 01m 16.6s	0.56%
ro	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	793h 15m 31.4s	0.56%
sk	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	788h 41m 35.1s	0.56%
sv	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	788h 48m 58.1s	0.56%
bg	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	772h 21m 20.5s	0.55%
cs	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	797h 43m 44.8s	0.57%
da	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	777h 27m 20.2s	0.55%
el	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	789h 53m 22.5s	0.56%
fi	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	787h 01m 03.0s	0.56%
hr	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	790h 43m 21.3s	0.56%
hu	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	783h 41m 55.2s	0.56%
lt	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	778h 44m 22.9s	0.55%
mt	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	778h 47m 18.4s	0.55%
pl	100,002	0.19%	0	0.00%	0	0.00%	100,002	0.18%	794h 50m 00.2s	0.56%
ko	92,182	0.17%	0	0.00%	0	0.00%	92,182	0.17%	217h 09m 58.0s	0.15%
ca	54,173	0.10%	12,730	3.37%	12,730	3.55%	79,633	0.15%	119h 48m 09.2s	0.09%
ru	12,112	0.02%	6,300	1.67%	6,110	1.70%	24,522	0.05%	38h 40m 15.3s	0.03%
zh-CN	7,085	0.01%	4,898	1.29%	4,843	1.35%	16,826	0.03%	26h 37m 10.1s	0.02%
fa	4,347	0.01%	3,445	0.91%	3,445	0.96%	11,237	0.02%	14h 20m 32.7s	0.01%
tr	3,966	0.01%	1,629	0.43%	1,624	0.45%	7,219	0.01%	7h 52m 17.3s	0.01%
mn	2,018	0.00%	1,759	0.47%	1,761	0.49%	5,538	0.01%	8h 21m 37.1s	0.01%
ar	2,029	0.00%	1,695	0.45%	1,758	0.49%	5,482	0.01%	5h 35m 01.6s	0.00%
sv-SE	2,160	0.00%	1,595	0.42%	1,349	0.38%	5,104	0.01%	4h 24m 34.4s	0.00%
id	1,243	0.00%	844	0.22%	792	0.22%	2,879	0.01%	2h 58m 58.8s	0.00%
ta	1,358	0.00%	786	0.21%	384	0.11%	2,528	0.00%	3h 04m 21.3s	0.00%
cy	721	0.00%	690	0.18%	690	0.19%	2,101	0.00%	3h 01m 18.5s	0.00%

Tag / Source Distribution

Value	train samples	train %	test samples	test %	validation samples	validation %	Total samples	Total %	Total Duration	Total Dur %
Emilia	40,237,834	75.00%	0	0.00%	0	0.00%	40,237,834	73.98%	101,585h 04m 02.8s	72.21%
GigaSpeech	5,053,116	9.42%	0	0.00%	0	0.00%	5,053,116	9.29%	6,297h 24m 07.6s	4.48%
CoVoST	3,925,255	7.32%	327,848	86.67%	323,985	90.37%	4,577,088	8.42%	7,114h 56m 33.0s	5.06%
MOSEL	2,300,046	4.29%	0	0.00%	0	0.00%	2,300,046	4.23%	18,127h 01m 37.8s	12.89%
PeopleSpeech	1,501,271	2.80%	34,898	9.23%	18,622	5.19%	1,554,791	2.86%	5,987h 42m 22.4s	4.26%
LibriTTS	353,817	0.66%	9,955	2.63%	10,340	2.88%	374,112	0.69%	585h 37m 48.6s	0.42%
Librispeech	281,241	0.52%	5,559	1.47%	5,567	1.55%	292,367	0.54%	982h 18m 30.2s	0.70%

License Distribution

Value	train samples	train %	test samples	test %	validation samples	validation %	Total samples	Total %
CC-BY-NC-4.0	40,237,834	75.00%	0	0.00%	0	0.00%	40,237,834	73.98%
unknown	5,053,116	9.42%	0	0.00%	0	0.00%	5,053,116	9.29%
CC0	3,925,255	7.32%	327,848	86.67%	323,985	90.37%	4,577,088	8.42%
CC-BY-4.0	2,935,104	5.47%	15,514	4.10%	15,907	4.44%	2,966,525	5.45%
CC-BY;CC-BY-SA	1,501,271	2.80%	34,898	9.23%	18,622	5.19%	1,554,791	2.86%

unknown rows correspond to the GigaSpeech subset. GigaSpeech's code and manifest are licensed under Apache 2.0, but the audio is governed by the GigaSpeech Data User Agreement and the platform terms (YouTube, podcasts, audiobooks) of the underlying clips. Use is limited to non-commercial academic research; raw-audio redistribution is prohibited upstream.

Data quality notes

MOSEL durations

Per-segment durations for all 2,300,046 MOSEL rows are sourced from VoxPopuli's authoritative segment manifest (unlabelled_v2.tsv), keyed by (event_id, segment_no) parsed from the row's path. Coverage is 100% — no MOSEL row has a missing or zero duration. Median segment is 30.0 s; minimum is 14.95 s.

MOSEL audio and language-label caveats

MOSEL audio is hosted in a separate repository, meetween/mumospee_mosel. Two known caveats apply to MOSEL-tagged rows and are documented on that repository's data card:

• Audio coverage is incomplete for some languages (voxpopuli.download_audios did not finish for the full language set). Per-segment transcripts and durations in this dataset remain valid for the affected rows; the missing audio can be fetched from FBK-MT/mosel or VoxPopuli upstream.
• Per-segment language labels are inherited from VoxPopuli's session-level metadata and are not human-verified at the utterance level. A fastText lid.176 sanity check disagrees with the declared language on ~1.5% of MOSEL rows, mostly on close-language pairs and short transcripts; the original declared labels are kept as-is here.

Subsets other than MOSEL (CoVoST, GigaSpeech, Emilia, PeoplesSpeech, LibriSpeech, LibriTTS) carry the language labels supplied by their original curators and are not independently verified.

Notes

• train: 0 rows with unparseable duration
• test: 0 rows with unparseable duration
• validation: 0 rows with unparseable duration
• 252 CoVoST rows have duration as NaN (source MP3 bytes are unreadable upstream — see Changelog below).

Mumospee dataset structure

Each row in the metadata represents one audio sample with the following fields:

• path: the relative path of the audio file
• url: the link to download the parquet shard containing the audio
• type: the sample type (audio or video)
• duration: duration in seconds
• language: language of the audio
• transcript: transcript text
• tag: origin dataset (one of _TAGS above)
• split: train, test, or validation
• license: license governing this sample

Example row:

{
  "path": "3660-172183-0000.flac",
  "url": "https://huggingface.co/datasets/meetween/mumospee_librispeech/resolve/main/librispeech-parquet/dev-other.parquet",
  "type": "audio",
  "duration": 5.405,
  "language": "en",
  "transcript": "GERAINT AS HE HAD BEEN USED TO DO WHEN HE WAS AT ARTHUR'S COURT FREQUENTED TOURNAMENTS",
  "tag": "Librispeech",
  "split": "validation",
  "license": "CC-BY-4.0"
}

Intended Uses

This dataset is designed to enable SpeechLLM and other large language models to support language-neutral virtual meeting applications.

Data Sources

The release includes metadata and download URLs for the following publicly available datasets:

• CoVoST (specifically CoVoST 2 — full 36 language pairs since 2026-06-22)
• GigaSpeech
• PeopleSpeech
• LibriSpeech
• LibriTTS
• Emilia
• MOSEL

Example usage

# pip install datasets

from datasets import load_dataset

# ── Load all splits at once ───────────────────────────────────────────────────

dataset = load_dataset("meetween/mumospee")
print(dataset)
# DatasetDict({
#     train:      Dataset({features: [...], num_rows: ...})
#     test:       Dataset({features: [...], num_rows: ...})
#     validation: Dataset({features: [...], num_rows: ...})
# })

# ── Load a specific split ─────────────────────────────────────────────────────

train_data      = load_dataset("meetween/mumospee", split="train")
test_data       = load_dataset("meetween/mumospee", split="test")
validation_data = load_dataset("meetween/mumospee", split="validation")

License

The metadata is published under CC-BY-4.0. Each individual sample is governed by its own license, recorded per-row in the license column. Users must comply with the licensing terms of each underlying dataset.

Changelog

2026-06-22

• Extended CoVoST coverage to the full CoVoST 2 catalog (36 language pairs, up from 27), adding en_id, ja_en, nl_en, pt_en, ru_en, sv-SE_en, ta_en, tr_en, and zh-CN_en. New rows reference parquet files in meetween/mumospee_covost, which was extended in the same release.

2026-06-08

• Repackaged from a single 43 GB dataset.csv into 29 sharded Parquet files (27 train, 1 test, 1 validation). Streaming, columnar reads, and load_dataset(...) are substantially faster, and CSV parsing edge cases are eliminated.
• Recomputed all headline statistics directly from the released data. Earlier versions of this card cited duration totals from upstream documentation; the figures were re-derived from the actual rows in the parquet shards.
• Backfilled MOSEL duration for all 2,300,046 rows from VoxPopuli's segment manifest. Previously these rows carried no real per-segment duration (the ingest hard-coded "n/a", which a downstream cleanup step had converted to 0.0, silently zeroing out ~18,000 hours of MOSEL).
• Row repair pass on the metadata:
  • rows whose transcript contained unescaped commas (and so were split into >9 fields) were re-joined into the correct 9-column shape;
  • rows whose duration carried a unit suffix (e.g. "3.5 s") were stripped to a numeric value;
  • rows whose language was empty or longer than 5 characters were dropped;
  • exact duplicate rows were removed.
• Corrected license attribution. The license value for tag == Emilia rows is now reported as CC-BY-NC-4.0 to match the upstream Emilia license; an earlier version of this card had bucketed those rows under CC-BY-4.0 in the distribution table even though the per-row values were already correct.
• Corrected MOSEL language for 55,387 rows by overwriting with the VoxPopuli interpretation-channel code embedded in path (e.g. ..._bg_3 → bg).
• duration set to NaN for 252 CoVoST rows whose source MP3 bytes are unreadable upstream (mutagen and ffprobe both fail to parse them). These rows still carry valid path, transcript, and language; users requiring duration should filter with df["duration"].notna().
• Schema verification: 0 rows with empty language, 0 rows with len(language) > 5, 0 rows with duration == 0 (252 CoVoST rows are null/NaN — see above — and excluded from duration statistics).

The row schema (path, url, type, duration, language, transcript, tag, split, license), the _TAGS set, and the load_dataset(...) filters are unchanged.