TalTechNLP/voxlingua107_wds

VoxLingua107

VoxLingua107 is a speech dataset for training spoken language identification models. The dataset consists of short speech segments automatically extracted from YouTube videos and labeled according the language of the video title and description, with some post-processing steps to filter out false positives.

VoxLingua107 contains data for 107 languages. The total amount of speech in the training set is 6628 hours. The average amount of data per language is 62 hours. However, the real amount per language varies a lot. There is also a seperate development set containing 1609 speech segments from 33 languages, validated by at least two volunteers to really contain the given language.

For more information, see the paper Jörgen Valk, Tanel Alumäe. VoxLingua107: a Dataset for Spoken Language Recognition. Proc. SLT 2021.

Why

VoxLingua107 can be used for training spoken language recognition models that work well with real-world, varying speech data. You can try a demo system trained on this dataset here.

How

We extracted audio data from YouTube videos that are retrieved using language-specific search phrases (random phrases from Wikipedia of the particular language). If the language of the video title and description matched with the language of the search phrase, the audio in the video was deemed likely to be in that particular language. This allowed to collect large amounts of somewhat noisy data relatively cheaply. Speech/non-speech detection and speaker diarization was used to segment the videos into short sentence-like utterances. A data-driven post-filtering step was applied to remove clips that were very different from other clips in this language's dataset, and thus likely not in the given language. Due to the automatic data collection process, there are still clips in the dataset that are not in the given language or contain non-speech (around 2% overall), especially for some languages (like Welsh).

License and copyright

The VoxLingua107 dataset is distributed under the Creative Commons Attribution 4.0 International License. The copyright remains with the original owners of the video.

We also point out that the distribution of languages, accents, dialects, genders, races and societal factors in this dataset is not representative of the global population. Using this dataset for training and deploying models may thus introduce unintended biases.

Notice and take down policy

Notice: Should you consider that our data contains material that is owned by you and should therefore not be reproduced here, please:

• Clearly identify yourself, with detailed contact data such as an address, telephone number or email address at which you can be contacted.
• Clearly identify the copyrighted work claimed to be infringed.
• Clearly identify the material that is claimed to be infringing and information reasonably sufficient to allow us to locate the material.
• Send the request to Tanel Alumäe

Take down: We will comply to legitimate requests by removing the affected sources from the corpus.

Languages and sizes

Language Code	Language Name	Hours
ab	Abkhazian	10
af	Afrikaans	108
am	Amharic	81
ar	Arabic	59
as	Assamese	155
az	Azerbaijani	58
ba	Bashkir	58
be	Belarusian	133
bg	Bulgarian	50
bn	Bengali	55
bo	Tibetan	101
br	Breton	44
bs	Bosnian	105
ca	Catalan	88
ceb	Cebuano	6
cs	Czech	67
cy	Welsh	76
da	Danish	28
de	German	39
el	Greek	66
en	English	49
eo	Esperanto	10
es	Spanish	39
et	Estonian	38
eu	Basque	29
fa	Persian	56
fi	Finnish	33
fo	Faroese	67
fr	French	67
gl	Galician	72
gn	Guarani	2
gu	Gujarati	46
gv	Manx	4
ha	Hausa	106
haw	Hawaiian	12
hi	Hindi	81
hr	Croatian	118
ht	Haitian	96
hu	Hungarian	73
hy	Armenian	69
ia	Interlingua	3
id	Indonesian	40
is	Icelandic	92
it	Italian	51
iw	Hebrew	96
ja	Japanese	56
jw	Javanese	53
ka	Georgian	98
kk	Kazakh	78
km	Central Khmer	41
kn	Kannada	46
ko	Korean	77
la	Latin	67
lb	Luxembourgish	75
ln	Lingala	90
lo	Lao	42
lt	Lithuanian	82
lv	Latvian	42
mg	Malagasy	109
mi	Maori	34
mk	Macedonian	112
ml	Malayalam	47
mn	Mongolian	71
mr	Marathi	85
ms	Malay	83
mt	Maltese	66
my	Burmese	41
ne	Nepali	72
nl	Dutch	40
nn	Norwegian Nynorsk	57
no	Norwegian	107
oc	Occitan	15
pa	Panjabi	54
pl	Polish	80
ps	Pushto	47
pt	Portuguese	64
ro	Romanian	65
ru	Russian	73
sa	Sanskrit	15
sco	Scots	3
sd	Sindhi	84
si	Sinhala	67
sk	Slovak	40
sl	Slovenian	121
sn	Shona	30
so	Somali	103
sq	Albanian	71
sr	Serbian	50
su	Sundanese	64
sv	Swedish	34
sw	Swahili	64
ta	Tamil	51
te	Telugu	77
tg	Tajik	64
th	Thai	61
tk	Turkmen	85
tl	Tagalog	93
tr	Turkish	59
tt	Tatar	103
uk	Ukrainian	52
ur	Urdu	42
uz	Uzbek	45
vi	Vietnamese	64
war	Waray	11
yi	Yiddish	46
yo	Yoruba	94
zh	Mandarin Chinese	44

Usage

Although webdataset can be used in a streaming fashion, it is recommended to first make a local copy fo the dataset using git clone.

git lfs install
git clone [email protected]:datasets/TalTechNLP/voxlingua107_wds

Then you can use the Python webdataset library to create an iterable dataset out of it:

import webdataset as wds
import random
train_files = glob.glob("voxlingua107_wds/train/**/*.tar") # you can also limit the training data to selected languages    
dev_files = glob.glob("voxlingua107_wds/dev/*.tar")

random.shuffle(train_files)

def mapper(sample):
    # "audio" field represents 16 kHz raw audio
    return {"audio": sample[0], "lang": sample[1]["lang"]}
    
# Since each shard contains 500 samples for a single language, it is good to use a reasonably large buffer size to get nicely shuffled samples
buffer_size = 100000   
dataset = wds.WebDataset(train_urls, shardshuffle=2000).shuffle(buffer_size, initial=buffer_size).decode(wds.torch_audio).to_tuple("wav","json").map(mapper)
dev_dataset = wds.WebDataset(dev_urls).decode(wds.torch_audio).to_tuple("wav","json").map(mapper)

train_iter = iter(dataset)
print(next(train_iter))
{'audio': (tensor([[-9.7656e-04, -8.5449e-04, -3.0518e-05,  ...,  2.7466e-03,
      3.7842e-03,  5.1880e-03]]), 16000), 'lang': 'kn'}

Citing

@inproceedings{valk2021slt,
  title={{VoxLingua107}: a Dataset for Spoken Language Recognition},
  author={J{\"o}rgen Valk and Tanel Alum{\"a}e},
  booktitle={Proc. IEEE SLT Workshop},
  year={2021},
}