Datasets:
[!4][RELEASE] MOSEL v1.0
Browse files- README.md +118 -0
- mosel-logo-transparent.png +0 -0
README.md
CHANGED
@@ -1,3 +1,121 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
license: cc-by-4.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
task_categories:
|
3 |
+
- automatic-speech-recognition
|
4 |
+
language:
|
5 |
+
- en
|
6 |
+
- bg
|
7 |
+
- hr
|
8 |
+
- cs
|
9 |
+
- da
|
10 |
+
- nl
|
11 |
+
- et
|
12 |
+
- fi
|
13 |
+
- fr
|
14 |
+
- de
|
15 |
+
- el
|
16 |
+
- hu
|
17 |
+
- ga
|
18 |
+
- it
|
19 |
+
- lv
|
20 |
+
- lt
|
21 |
+
- mt
|
22 |
+
- pl
|
23 |
+
- pt
|
24 |
+
- ro
|
25 |
+
- sk
|
26 |
+
- sl
|
27 |
+
- es
|
28 |
+
- sv
|
29 |
+
pretty_name: MOSEL
|
30 |
license: cc-by-4.0
|
31 |
---
|
32 |
+
|
33 |
+
<img src="./mosel-logo-transparent.png" align="center" width="100%">
|
34 |
+
|
35 |
+
### Dataset Description, Collection, and Source
|
36 |
+
|
37 |
+
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. We collect data by surveying labeled and unlabeled speech corpora under open-source compliant licenses.
|
38 |
+
In particular, MOSEL includes the automatic transcripts of 441k hours of unlabeled speech from VoxPopuli and LibriLight. The data is transcribed using [Whisper large v3](https://huggingface.co/openai/whisper-large-v3).
|
39 |
+
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.
|
40 |
+
|
41 |
+
- **Curated by:** Marco Gaido, Sara Papi, Luisa Bentivogli, Alessio Brutti, Mauro Cettolo, Roberto Gretter, Marco Matassoni, Mohamed Nabih, and Matteo Negri
|
42 |
+
- **Funded by:** FAIR, Meetween, and CINECA
|
43 |
+
- **Shared by:** Fondazione Bruno Kessler
|
44 |
+
|
45 |
+
### License
|
46 |
+
- CC-BY-4.0
|
47 |
+
|
48 |
+
### Dataset Sources
|
49 |
+
|
50 |
+
- **Collection Repository:** [MOSEL](https://github.com/hlt-mt/mosel)
|
51 |
+
- **Paper:** [MOSEL: 950,000 Hours of Speech Data for Open-Source Speech Foundation Model Training on EU Languages](https://arxiv.org/)
|
52 |
+
|
53 |
+
## Dataset Structure
|
54 |
+
|
55 |
+
### Data Config
|
56 |
+
The dataset is split into folders corresponding to the languages using the [2-letters ISO codes](https://en.wikipedia.org/wiki/List_of_ISO_639_language_codes), one for each language. Within each folder, a split for each psuedo-labeled dataset is provided.
|
57 |
+
|
58 |
+
### Data Field
|
59 |
+
`id`: alphanumeric identifier for the segment
|
60 |
+
|
61 |
+
`language`: extended language (e.g., "english")
|
62 |
+
|
63 |
+
`text`: the content of the psuedo label
|
64 |
+
|
65 |
+
`hall_repeated_ngrams`: True/False - indicates the repetition of an *n*-gram in `text` for a minimum number of times; for *n* in 1 to 2, the threshold is 4, for *n* in 3 to 5, it is 3
|
66 |
+
|
67 |
+
`hall_long_word`: True/False - indicates the presence of a word of at least 40 characters in `text`
|
68 |
+
|
69 |
+
`hall_frequent_single_word`: True/False - indicates that `text` consists of only one word which is the most frequent inside the whole text
|
70 |
+
|
71 |
+
## Dataset Statistics (in hours)
|
72 |
+
|
73 |
+
| Language (LangID) | Labeled | Unlabeled | Total |
|
74 |
+
|--------|--------|--------|-------|
|
75 |
+
| Bulgarian (bg) | 111 | 17609 | 17720 |
|
76 |
+
| Croatian (hr) | 55 | 8106 | 8161 |
|
77 |
+
| Czech (cs) | 591 | 18705 | 19296 |
|
78 |
+
| Danish (da) | 20 | 13600 | 13620 |
|
79 |
+
| Dutch (nl) | 3395 | 19014 | 22409 |
|
80 |
+
| English (en) | 437239 | 84704 | 521943|
|
81 |
+
| Estonian (et) | 60 | 10604 | 10664 |
|
82 |
+
| Finnish (fi) | 64 | 14200 | 14264 |
|
83 |
+
| French (fr) | 26984 | 22896 | 49880 |
|
84 |
+
| German (de) | 9236 | 23228 | 32464 |
|
85 |
+
| Greek (el) | 35 | 17703 | 17738 |
|
86 |
+
| Hungarian (hu) | 189 | 17701 | 17890 |
|
87 |
+
| Irish (ga) | 17 | 0 | 17 |
|
88 |
+
| Italian (it) | 3756 | 21933 | 25689 |
|
89 |
+
| Latvian (lv) | 173 | 13100 | 13273 |
|
90 |
+
| Lithuanian (lt) | 36 | 14400 | 14436 |
|
91 |
+
| Maltese (mt) | 19 | 9100 | 9119 |
|
92 |
+
| Polish (pl) | 510 | 21207 | 21717 |
|
93 |
+
| Portuguese (pt) | 5492 | 17526 | 23018 |
|
94 |
+
| Romanian (ro) | 121 | 17906 | 18021 |
|
95 |
+
| Slovak (sk) | 61 | 12100 | 12161 |
|
96 |
+
| Slovenian (sl) | 32 | 11300 | 11332 |
|
97 |
+
| Spanish (es) | 17471 | 21526 | 38997 |
|
98 |
+
| Swedish (sv) | 58 | 16300 | 16358 |
|
99 |
+
| Total | 505725 | 444467 | 950192|
|
100 |
+
|
101 |
+
|
102 |
+
## Dataset Creation
|
103 |
+
To reproduce the dataset creation, please refer to the [MOSEL README in the fbk-llm](https://github.com/hlt-mt/fbk-llm) repository.
|
104 |
+
|
105 |
+
|
106 |
+
## Citation
|
107 |
+
Release 1.0:
|
108 |
+
```
|
109 |
+
@inproceedings{mosel,
|
110 |
+
title = {{MOSEL: 950,000 Hours of Speech Data for Open-Source Speech Foundation Model Training on EU Languages}},
|
111 |
+
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},
|
112 |
+
booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
|
113 |
+
month = nov,
|
114 |
+
year = "2024",
|
115 |
+
address = "Miami, United States",
|
116 |
+
publisher = "Association for Computational Linguistics",
|
117 |
+
}
|
118 |
+
```
|
119 |
+
|
120 |
+
## Dataset Card Contact
|
121 |
+
[@spapi](https://huggingface.co/spapi)
|
mosel-logo-transparent.png
ADDED