Multilingual Speech Recognition (ASR) Datasets
Use Common Voice for broad crowd-sourced language coverage with a fully open CC0 license, FLEURS when you need a parallel, consistently structured evaluation set across 102 languages, Multilingual LibriSpeech for large-volume audiobook speech in 8 major languages, and VoxPopuli for high-volume European-parliament speech with both labeled and unlabeled splits. Below is what each dataset actually contains, its license, and how to pick between them.
The short answer
For training a multilingual ASR model, Common Voice gives the widest volunteer-driven language coverage under a fully open CC0 license; Multilingual LibriSpeech gives large per-language volume for 8 major languages under CC BY 4.0; VoxPopuli adds hundreds of thousands of hours of unlabeled speech alongside smaller labeled sets across 23 European languages. For evaluation specifically, FLEURS is the standard choice because it is a parallel dataset — the same sentences recorded across 102 languages — which makes cross-language comparisons fair.
Broad language coverage
Mozilla Common Voice
Common Voice is Mozilla's open, crowd-sourced speech dataset, built by volunteers worldwide who record themselves reading prompted sentences and other volunteers validate the recordings for accuracy. Language coverage is broad — over 100 languages have some data — but volume and validated-clip quality vary enormously between well-resourced languages with large volunteer communities and low-resource languages with only a small amount of validated audio. Audio clips are released under CC0 (public domain), which permits unrestricted commercial use. Get it from commonvoice.mozilla.org/en/datasets.
FLEURS
FLEURS (Few-shot Learning Evaluation of Universal Representations of Speech), from Google Research, is a parallel evaluation dataset covering 102 languages, built by recording the same set of sentences (drawn from the FLoRes machine-translation dataset) across every language, so every language has comparable content and structure. That parallel design makes FLEURS particularly well suited to benchmarking ASR and speech-translation systems across many languages on equal footing, rather than training a production model from scratch, since its per-language volume is modest compared to Common Voice or MLS. Get it from Hugging Face: google/fleurs.
Large-volume single-domain speech
Multilingual LibriSpeech (MLS)
MLS extends the original LibriSpeech recipe to 8 languages by deriving read-audiobook speech and aligned transcripts from LibriVox recordings. Volume is substantial — the English split alone totals roughly 2.4TB uncompressed (651GB compressed) — while other languages range from a few gigabytes (Polish) up to over a hundred gigabytes (German), each split into train, development and test sets. Because it is derived from read audiobooks rather than spontaneous or noisy speech, it works best for building a strong acoustic baseline rather than testing robustness to conversational or noisy audio. It is CC BY 4.0 licensed, permitting commercial use with attribution. Get it from openslr.org/94.
VoxPopuli
VoxPopuli is built from European Parliament recordings spanning 2009–2020, giving it an unusually large amount of raw speech: about 400,000 hours of unlabeled speech across 23 languages, plus roughly 1,800 hours of transcribed speech for 16 of those languages and additional speech-to-speech interpretation data. Its scale makes it a strong source for self-supervised pretraining (the project's own wav2vec 2.0 baselines were trained on it), while the smaller transcribed subset supports supervised fine-tuning. The raw audio and transcripts are CC0, but the accompanying code and pretrained models are released under CC BY-NC 4.0 (non-commercial) — check which asset you actually need before assuming the whole release is commercially usable. Get it from github.com/facebookresearch/voxpopuli.
Datasets side by side
| Dataset | Languages | Scale | License |
|---|---|---|---|
| Common Voice | 100+ (variable volume per language) | Grows continuously via crowd-sourcing | CC0 (audio) |
| FLEURS | 102, parallel content | Small-to-moderate per language, evaluation-focused | Check dataset page |
| Multilingual LibriSpeech | 8 (audiobook speech) | 2.4TB English; 6GB–115GB per other language | CC BY 4.0 |
| VoxPopuli | 23 (European Parliament) | 400K hrs unlabeled; 1.8K hrs transcribed | CC0 (data); CC BY-NC 4.0 (code/models) |
How to choose
If you need broad language coverage and the freedom to redistribute commercially, start with Common Voice, but check per-language validated-hours counts first since coverage is uneven. If you are evaluating a multilingual model and need a fair, apples-to-apples comparison across languages, use FLEURS rather than mixing datasets with different recording conditions. If your target languages fall within MLS's 8 supported languages and you need a large, clean training corpus, MLS gives you more hours per language than Common Voice for those languages. If you are pretraining a self-supervised speech model and want scale above all else, VoxPopuli's 400K hours of unlabeled speech is hard to match, but confirm which license applies to the specific asset (raw audio versus code/models) you plan to use.
FAQ
Which multilingual ASR dataset has the broadest language coverage?
FLEURS covers 102 languages with parallel evaluation speech, making it the broadest standard benchmark for many-language ASR and speech-to-text evaluation. Common Voice has broad crowd-sourced coverage across roughly 100+ languages as well, though the amount of validated data per language varies widely. VoxPopuli covers 23 European languages but with a much larger volume of hours per language.
Can I use Common Voice and Multilingual LibriSpeech commercially?
Common Voice audio clips are released under CC0, which permits commercial use without attribution. Multilingual LibriSpeech is CC BY 4.0, which permits commercial use with attribution. VoxPopuli's raw audio and transcripts are CC0, but its code and pretrained models are CC BY-NC 4.0 (non-commercial) — check which asset you need before assuming the whole release is commercially usable.
What is the difference between Common Voice and FLEURS for evaluation?
Common Voice is built from open crowd-sourced volunteer recordings of arbitrary sentences, so its size and audio quality vary by language, skewing toward languages with an active volunteer community. FLEURS is a parallel n-way evaluation set — the same sentences recorded across all 102 languages — which makes it better suited to comparing ASR or speech-translation performance consistently across languages, even though its per-language volume is much smaller than Common Voice's larger languages.
Still cannot find the right speech data? Leave it to us
If your target language or domain — a dialect, an industry-specific vocabulary, a low-resource language — is not well covered by these datasets, tell us your task goal and the conditions it must meet. DeepSData can run a real availability search and verify sources for you, checking reputable data platforms and judging matches and gaps against each requirement, with every source link kept traceable.
New accounts get 3 free searches; every result comes with verifiable source links. See also curated datasets.