meetween/mumospee_mosel

Dataset Summary

This dataset is a modified version of the MOSEL and VoxPopuli corpus, converted into parquet format to facilitate optimized I/O operations in high-performance and distributed computing environments. The MOSEL corpus is a multilingual dataset collection including up to 950K hours of open-source speech recordings covering the 24 official languages of the European Union. MOSEL includes the automatic transcripts of 441k hours of unlabeled speech from VoxPopuli and LibriLight. The data is transcribed using Whisper large v3. Whisper is released under the OS Apache 2.0 License which allows releasing the generated content under any license. Since LibriLight, differently from VoxPopuli, contains segments longer than Whisper's maximum duration limit of 30sec, we split them into chunks of up to 30sec.

This release covers the VoxPopuli PLENARY subset of MOSEL only. COMMITTEE, SPECIAL, and DEBATE sessions are not included; both the audio and the segment metadata are restricted to PLENARY content so that every metadata row points to a session present in this repository.

---

Source Data

• Original Dataset: MOSEL, VoxPopuli
• License: This derived dataset is shared under the same license, with modifications only to format for efficiency.

Modifications

• Data Format: Converted to parquet format to enhance I/O performance for distributed training, reducing latency during data loading and retrieval.
• Efficiency Optimization: Restructured for reduced storage footprint and faster I/O on high-performance clusters by leveraging parquet's efficient compression and columnar storage.
• Scope: Restricted to VoxPopuli PLENARY sessions across 23 EU languages, so the audio and metadata folders stay aligned.

Dataset Structure

• File Format: Parquet files.
• Languages: 23 EU languages (bg, cs, da, de, el, en, es, et, fi, fr, hr, hu, it, lt, lv, mt, nl, pl, pt, ro, sk, sl, sv).
• Audio Sampling Rate: 16 kHz mono, OGG/Vorbis encoded.

The repository contains two top-level folders:

mosel-parquet/ — whole-session audio

Sharded by (language, year) at mosel-parquet/{lang}_{year}.parquet. Each row is one whole-session OGG file.

column	type	description
file_name	string	e.g. 20200113-0900-PLENARY_en.ogg
sample_rate	int64	16000
duration_ms	int64	whole-session duration in milliseconds
samples	binary	raw OGG bytes (Vorbis-encoded, 16 kHz mono)

Segment-level slicing is performed downstream using offsets from FBK-MT/mosel/<lang>/voxpopuli.tsv (or the VoxPopuli unlabelled_v2 manifest).

metadata/ — per-segment metadata

One file per language at metadata/mosel_<lang>.parquet. Each row is one MOSEL segment.

column	type	description
path	string	segment id, e.g. 20160118-0900-PLENARY-12_en_3
url	string	URL of the mosel-parquet/ shard that holds this segment's session audio
type	string	constant "audio"
duration	float64	segment duration in seconds
language	string	source / interpretation-channel language code (parsed from path)
transcript	string	Whisper-large-v3 transcript
tag	string	constant "MOSEL"
split	string	constant "train"
license	string	constant "CC-BY-4.0"

Totals: 12,679,032 rows across 23 languages, ~100,317 hours of audio.

Statistics

Overall

• 23 languages, 12,679,032 segments, ~100,317 hours of audio.
• 66 (lang, year) audio shards covering 2009–2020 (mean ~918 sessions per shard).
• Average segment duration: 28.5 s (close to Whisper's 30 s cap; the slightly lower mean reflects shorter trailing chunks per session).

Per language

lang	rows	hours	sessions	avg segment (s)
bg	553,022	4,325.3	7,676	28.16
cs	565,495	4,517.1	7,751	28.76
da	553,516	4,322.7	7,637	28.11
de	563,210	4,512.5	7,687	28.84
el	556,641	4,398.1	7,659	28.44
en	566,970	4,531.6	7,680	28.77
es	565,547	4,498.4	7,748	28.63
et	556,894	4,353.9	7,654	28.15
fi	563,305	4,472.4	7,663	28.58
fr	566,970	4,522.5	7,705	28.72
hr	339,551	2,683.6	4,058	28.45
hu	559,069	4,390.0	7,663	28.27
it	569,156	4,558.5	7,754	28.83
lt	551,298	4,297.9	7,609	28.07
lv	561,451	4,435.7	7,688	28.44
mt	556,919	4,361.5	7,667	28.19
nl	564,295	4,489.7	7,703	28.64
pl	567,309	4,517.1	7,677	28.66
pt	562,416	4,432.9	7,699	28.37
ro	563,319	4,481.1	7,679	28.64
sk	545,302	4,315.2	7,476	28.49
sl	559,384	4,387.3	7,636	28.24
sv	567,993	4,511.9	7,676	28.60
total	12,679,032	100,317.0	—	—

Languages are very evenly distributed (~4.4 % each) except hr, which is smaller (2.7 %) because MOSEL VoxPopuli has less Croatian content overall.

Per year

year	rows	hours	% of total
2009	949,131	7,493.4	7.49 %
2010	1,139,862	9,028.4	8.99 %
2011	1,207,413	9,620.6	9.52 %
2012	1,020,548	8,099.1	8.05 %
2013	1,182,756	9,348.9	9.33 %
2014	959,228	7,624.5	7.57 %
2015	1,334,005	10,562.3	10.52 %
2016	929,677	7,367.1	7.33 %
2017	1,206,722	9,485.6	9.52 %
2018	1,280,965	10,035.4	10.10 %
2019	953,620	7,518.7	7.52 %
2020	515,105	4,132.9	4.06 %

2020 is a partial year (data collection ends mid-year); other years are roughly proportional to plenary session count.

Usage

This dataset is ideal for use in large-scale multilingual speech-to-text translation tasks, especially in distributed and high-performance computing environments. The parquet format enhances usability by minimizing I/O overhead, making it well-suited for high-throughput training.

Attribution

This dataset is based on the original MOSEL and VoxPopuli dataset, with modifications for I/O optimization by converting to parquet format. Please cite the original CoVoST dataset in any publications or projects using this dataset.

Changelog

Version 1.1 (2026-05-26)

• Rebuilt 8 unopenable mosel-parquet/ shards (de_2017, fi_2017, mt_2017, mt_2018, pt_2018, ro_2017, ro_2018, sl_2017) and partially rebuilt bg_2009.parquet from the project's S3 raw-audio mirror.
• Replaced the per-language CSV metadata with PLENARY-only parquet metadata, sourced from FBK-MT/mosel and joined against the VoxPopuli unlabelled_v2 manifest for real per-segment durations.
• Fixed the language column to use the source / interpretation-channel code parsed from path instead of the Whisper langid output.
• Added cs, da, de metadata parquets for the first time.

Citation

Release 1.0:

@inproceedings{mosel,
  title = {{MOSEL: 950,000 Hours of Speech Data for Open-Source Speech Foundation Model Training on EU Languages}},
  author = {Marco Gaido and Sara Papi and Luisa Bentivogli and Alessio Brutti and Mauro Cettolo and Roberto Gretter and Marco Matassoni and Mohamed Nabihand Matteo Negri},
  booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
  month = nov,
  year = "2024",
  address = "Miami, United States",
  publisher = "Association for Computational Linguistics",
}