VoxPopuli

https://aclanthology.org/2021.acl-long.80

A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation.

Overview

VoxPopuli provides

• 400K hours of unlabelled speech data for 23 languages
• 1.8K hours of transcribed speech data for 16 languages
• 17.3K hours of speech-to-speech interpretation data for 15x15 directions

The raw data is collected from 2009-2020 European Parliament event recordings. We acknowledge the European Parliament for creating and sharing these materials.

Detailed statistics

Unlabelled and transcribed data

Language	Code	Unlabelled Hours (v1/v2)	Transcribed Hours	Transcribed Speakers	Transcribed Tokens	LM Tokens
English	En	4.5K/24.1K	543	1313	4.8M	60.1M
German	De	4.5K/23.2K	282	531	2.3M	50.0M
French	Fr	4.5K/22.8K	211	534	2.1M	58.6M
Spanish	Es	4.4K/21.4K	166	305	1.6M	57.4M
Polish	Pl	4.5K/21.2K	111	282	802K	13.6M
Italian	It	4.6K/21.9K	91	306	757K	52.1M
Romanian	Ro	4.5K/17.9K	89	164	739K	10.3M
Hungarian	Hu	4.4K/17.7K	63	143	431K	13.0M
Czech	Cs	4.5K/18.7K	62	138	461K	13.5M
Dutch	Nl	4.5K/19.0K	53	221	488K	54.6M
Finnish	Fi	4.4K/14.2K	27	84	160K	34.5M
Croatian	Hr	2.7K/8.1K	43	83	337K	285K
Slovak	Sk	4.4K/12.1K	35	96	270K	13.3M
Slovene	Sl	4.4K/11.3K	10	45	76K	12.6M
Estonian	Et	4.3K/10.6K	3	29	18K	11.3M
Lithuanian	Lt	4.3K/14.4K	2	21	10K	11.5M
Portuguese	Pt	4.4K/17.5K	-	-	-	-
Bulgarian	Bg	4.3K/17.6K	-	-	-	-
Greek	El	4.4K/17.7K	-	-	-	-
Latvian	Lv	4.4K/13.1K	-	-	-	-
Maltese	Mt	4.4K/9.1K	-	-	-	-
Swedish	Sv	4.5K/16.3K	-	-	-	-
Danish	Da	4.3K/13.6K	-	-	-	-
Total		100K/384K	1791	4295	15M	467M

Speech-to-speech interpretation data

Source/Target	En	De	Fr	Es	Pl	It	Ro	Hu	Cs	Nl	Fi	Sk	Sl	Lt	Da	Total
En	-	463	427	441	432	461	457	382	427	400	442	433	434	398	370	6.0K
De	187	-	196	204	214	217	198	205	214	196	217	208	218	164	179	2.8K
Fr	169	187	-	187	172	197	195	144	170	158	168	168	156	139	134	2.3K
Es	130	138	135	-	118	148	128	93	118	115	124	114	108	83	86	1.6K
Pl	68	66	54	55	-	67	55	43	67	42	55	62	57	50	34	775
It	69	77	76	79	72	-	75	61	68	64	71	66	70	53	60	961
Ro	60	59	59	58	49	61	-	38	50	43	48	50	46	38	29	688
Hu	30	38	25	27	29	30	27	-	27	20	31	29	26	21	18	378
Cs	39	35	29	30	36	32	31	23	-	23	29	55	29	25	18	434
Nl	31	43	35	29	27	38	24	25	25	-	32	25	23	19	25	401
Fi	15	18	15	13	13	13	13	12	13	11	-	14	12	11	9	182
Hr	31	27	27	24	27	28	24	22	24	22	24	26	37	21	20	384
Sk	21	22	14	16	19	16	16	14	32	13	16	-	17	13	10	239
Sl	6	6	4	5	5	6	5	4	5	4	5	6	-	4	3	68
Lt	1	1	1	1	1	1	1	1	1	1	1	1	1	-	0	13
Total	857	1.2K	1.1K	1.2K	1.2K	1.3K	1.2K	1.1K	1.2K	1.1K	1.3K	1.3K	1.2K	1.0K	995	17.3K

Getting Data

We provide raw audios as well as scripts to segment and align them with transcription/interpretation. The output format is Ogg Vorbis (16000Hz, 16-bit, mono-channel), which is supported by common libraries such as libsndfile and libsox (they have Python frontends by soundfile, torchaudio, etc.).

As the first step, clone this repo for the processing scripts

git clone https://github.com/facebookresearch/voxpopuli.git

and install required PyPI packages:

pip install -r requirements.txt

Unlabelled Data

First, download raw audios via

python -m voxpopuli.download_audios --root [ROOT] --subset [SUBSET]

which saves audios to ${ROOT}/raw_audios/[language]/[year]/[recording_id].ogg.

SUBSET specifies the data subset to download:

--subset	# Languages	Hours	Years	Size
en, de, fr, es, pl, it, ro, hu, cs, nl, fi, hr, sk, sl, et, lt, pt, bg, el, lv, mt, sv or da	1	2.7K-4.6K	2009-2020	44G-75G
en_v2, de_v2, fr_v2, es_v2, pl_v2, it_v2, ro_v2, hu_v2, cs_v2, nl_v2, fi_v2, hr_v2, sk_v2, sl_v2, et_v2, lt_v2, pt_v2, bg_v2, el_v2, lv_v2, mt_v2, sv_v2 or da_v2	1	8.1K-24.1K	2009-2020	130G-385G
10k	23	10K	2019-2020	170G
100k	23	100K	2009-2020	1.7T
400k	23	400K	2009-2020	6.4T

Then, segment these audios via

python -m voxpopuli.get_unlabelled_data --root [ROOT] --subset [SUBSET]

which outputs to ${ROOT}/unlabelled_data/[language]/[year]/[segment_id].ogg

Transcribed (ASR) Data

First, download raw audios via

python -m voxpopuli.download_audios --root [ROOT] --subset asr

which saves audios to ${ROOT}/raw_audios/original/[year]/[recording_id].ogg.

Then, segment these audios and align them with transcripts via

python -m voxpopuli.get_asr_data --root [ROOT] --lang [LANGUAGE]

which outputs

• audios ${ROOT}/transcribed_data/[language]/[year]/[segment_id].ogg
• per-split manifest (ID, transcript, speaker ID) ${ROOT}/transcribed_data/[language]/asr_[split].tsv

Speech-to-Speech Interpretation Data

First, follow the instructions above to set up ASR data (source audios and transcripts).

Then, download target audios via

python -m voxpopuli.download_audios --root [ROOT] --subset [TARGET_LANGUAGE]

which saves audios to ${ROOT}/raw_audios/[target_language]/[year]/[recording_id].ogg.

Finally, segment these audios and match them with source ones via

python -m voxpopuli.get_s2s_data --root [ROOT] --source-lang [SOURCE_LANGUAGE] --target-lang [TARGET_LANGUAGE]

which outputs

• target audios ${ROOT}/transcribed_data/[language]/[target_language]/[year]/[segment_id].ogg
• manifest (source ID, transcript, speaker ID, target ID) ${ROOT}/transcribed_data/[language]/[target_language]/s2s.tsv

We also human-transcribe part of the target audios (for English, French and Spanish only) to allow more accurate alignments. To use them instead of machine transcriptions in the alignments, add --use-annotated-target to the command line.

Language Modeling (LM) Data

We combine VoxPopuli transcripts and text data from Europarl for LM training.

Download VoxPopuli and Europarl text data, process the raw text and generate the vocabulary via

python -m voxpopuli.get_lm_data --root [ROOT] --lang [LANGUAGE]

which outputs

• sentences ${ROOT}/lm_data/[language]/sentences.txt
• vocabulary ${ROOT}/lm_data/[language]/vocabulary.txt

To train an n-gram LM with KenLM, run

${KENLM_PATH}/lmplz -o ${n} --limit_vocab_file [OUT_VOCAB_FILE] < [OUT_TEXT_FILE] > ${n}gram_lm.arpa
${KENLM_PATH}/build_binary ${n}gram_lm.arpa ${n}gram_lm.bin

Pre-trained Models

wav2vec 2.0

We provide pre-trained wav2vec 2.0 models (implemented in fairseq and wav2letter/flashlight) for downstream speech tasks:

Language(s)	Pre-training Hours	Base Model (95M)	Large Model (317M)
Es	4.4K	fairseq	fairseq
Fr	4.5K	fairseq	fairseq
It	4.6K	fairseq	fairseq
Nl	4.5K	fairseq	fairseq
Sv	4.5K	fairseq	fairseq
All 23 languages	10K	fairseq	fairseq
All 23 languages	100K	fairseq / wav2letter	fairseq

In our paper (Section 4.3.1), we evaluated these models on the Common Voice corpus in the normal setting and the few-shot phoneme recognition setting.

Wav2letter C++ implementation

A wav2letter implementation as well as a checkpoint pretrained on VoxPopuli 100k (base model) is also available in the Wav2letter respository.

The complete fine-tuned ASR baselines for this codebase shoulda come soon. The wav2letter implementation follows this paper.

ASR and LM

For the VoxPopuli ASR task, we provide Transformer baselines, fine-tuned wav2vec2 models (Base 10K) as well as n-gram LMs (trained with KenLM) and their lexicons:

Language	ASR (fairseq)	LM (kenLM)	Lexicon
Cs	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
De	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
En	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Es	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Et	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Fi	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Fr	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Hr	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Hu	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
It	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Lt	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Nl	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Pl	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Ro	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Sk	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon
Sl	baseline, fine-tuned wav2vec2	3-gram, 5-gram	lexicon

We also provide CoVoST 2 + EuroParl-ST ASR Transformer models that are self-trained on 3000h VoxPopuli unlabelled data.

Language	CoVoST 2 Test (WER)	EuroParl-ST Test (WER)	Model (fairseq)
De	17.3	21.4	s2t_transformer_l
Es	13.2	15.3	s2t_transformer_l
Fr	17.0	19.0	s2t_transformer_l

Please refer to the S2T examples for the use of Transformer model checkpoints.

Speech-to-Text Translation (ST)

We provide CoVoST 2 + EuroParl-ST ST Transformer models that are jointly trained with 400h VoxPopuli weakly labelled data.

Direction	CoVoST 2 Test (BLEU)	EuroParl-ST Test (BLEU)	Model (fairseq)
De-En	23.4	24.4	s2t_transformer_l
Es-En	29.7	28.4	s2t_transformer_l
Fr-En	30.3	31.1	s2t_transformer_l

Please refer to the S2T examples for the use of these checkpoints.

What's New

• 2021-07-26: New unlabelled data (additional 300K hours) released.
• 2021-03-03: VoxPopuli released.

License

	License
VoxPopuli Data	CC0 (see also European Parliament's legal notice for the raw data)
LM Data	(Please check out the Europarl website for the Europarl portion)
Pre-trained Models	CC BY-NC 4.0
Code	CC BY-NC 4.0

Contact

Changhan Wang ([email protected]), Morgane Rivière ([email protected]), Ann Lee ([email protected])

Citation

@inproceedings{wang-etal-2021-voxpopuli,
    title = "{V}ox{P}opuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation",
    author = "Wang, Changhan  and
      Riviere, Morgane  and
      Lee, Ann  and
      Wu, Anne  and
      Talnikar, Chaitanya  and
      Haziza, Daniel  and
      Williamson, Mary  and
      Pino, Juan  and
      Dupoux, Emmanuel",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-long.80",
    pages = "993--1003",
}