Granary-v2: Multilingual Speech Recognition, Translation and Instruction Data

Granary-v2 is NVIDIA's second release of generated labels and instruction pairs for multilingual speech. It re-processes a broad set of source corpora through an entirely new labelling pipeline that produces not just transcripts and translations, but inverse text normalization, captioning, code-switching and speech-QA targets, plus ready-to-train instruction pairs.

The pipeline source is already public. All of it lives in NeMo Curator. The full end-to-end runs and the per-stage scripts will follow soon.

Overview

• 48 languages across 16 source corpora
• Labels for 463,062 hours of source speech
• approximately 165,300 hours flagged high-quality by the pipeline's own gate
• approximately 92,496 hours of X→English speech translation
• approximately 39,375 hours of English→X speech translation, each record fanned out to 31 targets at once (ar, bg, cs, da, de, el, es, et, fi, fr, he, hi, hr, hu, it, ja, ko, lt, lv, mt, nl, pl, pt, ro, ru, sk, sl, sv, th, uk, zh)
• Six text targets per utterance: normalized transcript, ITN, captioning, code-switched, speech-QA, and translation

Supported Languages

Arabic, Assamese, Bengali, Bulgarian, Chinese, Croatian, Czech, Danish, Dutch, English, Estonian, Filipino, Finnish, French, German, Greek, Gujarati, Hebrew, Hindi, Hungarian, Indonesian, Italian, Japanese, Kannada, Korean, Latvian, Lithuanian, Malayalam, Maltese, Marathi, Odia, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Vietnamese.

Hours by corpus

Corpus	Languages	Hours	High-quality h	AST (X→en) h	AST (en→X) h	Source corpus license
VoxPopuli (MOSEL labels) (voxpopuli)	23	380,711	123,192	84,999	8,990	CC0 / CC-BY-4.0 labels
LibriLight (librilight)	1	49,334	25,898	0	25,725	MIT
People's Speech (people_speech)	1	8,598	3,807	0	826	CC BY (clean partition)
Multilingual LibriSpeech (mls)	8	7,805	3,073	1,848	103	CC-BY-4.0
FalAR (falar)	1	4,715	1,605	1,602	—	CC-BY-4.0
Common Voice 26.0 (mcv26)	39	3,871	2,837	2,372	496	CC0-1.0
National Speech Corpus (nsc)	1	2,785	1,647	0	1,689	Singapore Open Data Licence
VoxPopuli (re-labelled) (voxpopuli_improved)	16	1,277	527	562	160	CC0 / CC-BY-4.0 labels
Golos (golos)	1	1,208	715	744	—	Custom (Salute Developers)
Europarl-ASR (europarl)	1	1,023	605	0	613	CC-BY-4.0
LibriSpeech-PC (librispeech_pc)	1	953	953	0	686	CC-BY-4.0
FLEURS (fleurs)	48	412	213	226	5	CC-BY-4.0
Dusha (dusha)	1	204	104	143	—	Custom (Salute Developers)
VCTK (vctk)	1	82	76	0	76	ODC-By 1.0
AMI Meeting Corpus (ami)	1	70	24	0	4	CC-BY-4.0
Google Speech Commands (google_speech_commands)	1	15	15	0	2	CC-BY-4.0
Total	48	463,062	~165,300	92,496	39,375

Totals describe only the manifests in this public release. ASR hours and translation into English are calculated from the reported corpus durations. High-quality and translation-out totals are approximate because the source values are rounded. Translation targets do not add audio hours.

The table lists source-corpus licenses. The released labels and instruction pairs are governed by OpenMDW-1.1, except for the people_speech subset, which is governed by CC BY-SA 4.0. See License for the subset mapping.

This repository contains label manifests only. Obtain audio from the linked source distributors under their applicable licenses and access terms.

🎧 Getting the audio

The commands below obtain source corpora from their distributors. Check the source licenses and access terms before downloading. Loading nvidia/Granary-v2 downloads label manifests only; obtain source audio separately.

This repository ships manifests only — no audio travels with it. Every audio_filepath is instead a self-describing identifier:

data_<lang>_<corpus>_<source path, with "/" replaced by "_">

So data_de_mcv26_cv-corpus-26.0-2026-06-12_de_clips_common_voice_de_21137340.opus is cv-corpus-26.0-2026-06-12/de/clips/common_voice_de_21137340.opus in the Common Voice 26.0 release. Fetch a corpus from its own distributor, then join on that suffix.

Source audio remains subject to its distributor's terms. See the Source corpus license column in Hours by corpus. See License for the terms that apply to the released labels and instruction pairs.

The four routes below are ordered by how much work the download takes. Open the one your corpus is in.

1 · One HF command people_speech · 8,598 h falar · 4,715 h fleurs · 412 h

2 · Direct download librilight · 49,334 h mls · 7,805 h golos · 1,208 h librispeech_pc · 953 h dusha · 204 h vctk · 82 h google_speech_commands · 15 h

3 · Corpus tooling voxpopuli · 380,711 h voxpopuli_improved · 1,277 h europarl · 1,023 h ami · 70 h

4 · Gated access mcv26 · 3,871 h nsc · 2,785 h

1 · One Hugging Face command

pip install -U "huggingface_hub[cli]" datasets

Corpus	Hours	How
People's Speech (people_speech)	8,598	hf download MLCommons/peoples_speech --repo-type dataset --include "clean/*" — we release labels for the government-origin subset of clean. The released manifests are governed by CC BY-SA 4.0. Obtain source audio under the distributor's terms.
FalAR (falar)	4,715	load_dataset("inesc-id/FalAR") — the audio is embedded in the Parquet shards, so load it rather than fetching files
FLEURS (fleurs)	412	load_dataset("google/fleurs", "ar_eg") — the config is <lang>_<region>

2 · Direct download

Corpus	Hours	Archive
LibriLight (librilight)	49,334	https://dl.fbaipublicfiles.com/librilight/data/large.tar (3.06 TB) — we use the large 60k-hour segmentation; small and medium sit alongside it in the repo
Multilingual LibriSpeech (mls)	7,805	https://dl.fbaipublicfiles.com/mls/mls_<lang>.tar.gz for english, german, dutch, french, spanish, italian, portuguese, polish — an _opus variant of each exists; full index at OpenSLR 94
Golos (golos)	1,208	https://cdn.chatwm.opensmodel.sberdevices.ru/golos/train_crowd{0..9}.tar (127 GB total) — we use the crowd partition, and crowd9 carries the manifests. Whole set in Opus: golos_opus.tar
LibriSpeech-PC (librispeech_pc)	953	The audio is plain LibriSpeech: https://www.openslr.org/resources/12/train-{clean-100,clean-360,other-500}.tar.gz. OpenSLR 145 adds only the punctuated and capitalised text
Dusha (dusha)	204	https://cdn.chatwm.opensmodel.sberdevices.ru/dusha/crowd.tar (30 GB) — crowd only; the podcast partition ships no audio for licensing reasons and we use none of it
VCTK (vctk)	82	https://datashare.ed.ac.uk/bitstream/handle/10283/3443/VCTK-Corpus-0.92.zip (11.7 GB)
Google Speech Commands (google_speech_commands)	15	http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz (2.4 GB)

3 · Needs the corpus's own tooling

VoxPopuli — the two VoxPopuli leaves come from the same repository but different subsets, so they take different commands. voxpopuli (380,711 h) is the v2 unlabelled release, carrying the MOSEL pseudo-labels that seeded ours (FBK-MT/mosel); voxpopuli_improved (1,277 h) is the transcribed ASR subset, re-labelled by us.

Mind the _v2 suffix. --subset bg is the v1 release and is roughly 4× smaller than the bg_v2 our manifests were built from.

git clone https://github.com/facebookresearch/voxpopuli && cd voxpopuli
pip install -r requirements.txt

# `voxpopuli` -> the v2 UNLABELLED subset, per language ("400k" fetches all 23 at once)
python -m voxpopuli.download_audios     --root ./data --subset bg_v2
python -m voxpopuli.get_unlabelled_data --root ./data --subset bg_v2

# `voxpopuli_improved` -> the transcribed ASR subset
python -m voxpopuli.download_audios --root ./data --subset asr
python -m voxpopuli.get_asr_data    --root ./data --lang cs

Europarl-ASR (europarl, 1,023 h) — clone the MLLP repository and run its fetch scripts:

git clone https://www.mllp.upv.es/git-pub/ggarces/Europarl-ASR.git

AMI (ami, 70 h) — download through the Lhotse recipe with the IHM (individual headset microphone) setting; every record we ship is an individual headset channel, so SDM/MDM audio will not align.

pip install lhotse && lhotse download ami --mic ihm ./ami

4 · Gated access

Corpus	Hours	What it takes
Common Voice 26.0 (mcv26)	3,871	Download the per-language cv-corpus-26.0-2026-06-12 archive from the Common Voice portal (email-gated); the mozilla-foundation/common_voice_* mirrors on the Hub lag the latest release
National Speech Corpus (nsc)	2,785	Request access from IMDA Singapore — there is no direct download

Repository Structure

nvidia/Granary-v2/
└── <language>/                 # ISO 639-1 codes
    └── <corpus>/
        └── <lang>_asr.jsonl.gz

One or more manifest files per (language, corpus). Large manifests can be split into shards.

Unlike Granary-v1, there is no separate <lang>_ast-en.jsonl: every translation travels inline in each record's translations field, alongside its quality-estimation score. The manifest files for each (language, corpus) carry both the ASR and the AST targets.

translations is keyed by target language, and the release runs in both directions. Non-English audio is translated into English, which is the AST (X→en) h column. English audio is translated the other way, into 31 targets at once (ar, bg, cs, da, de, el, es, et, fi, fr, he, hi, hr, hu, it, ja, ko, lt, lv, mt, nl, pl, pt, ro, ru, sk, sl, sv, th, uk, zh) — that is the AST (en→X) h column, and it is why corpora like LibriLight and VCTK show no X→English hours yet still carry translation.

Coverage is per-record, and a present key is not a usable translation: the pipeline writes a stub with an empty translation_raw where its quality estimator rejected the candidate. Test translation_raw for non-emptiness rather than assuming a key means a translation. Both table columns count distinct audio, once per record however many targets it carries, so neither double-counts and the two never sum past a corpus's hours.

🚀 Quick Start

These examples load the released label manifests. Obtain the source audio separately as described in Getting the audio. The people_speech records retain their CC BY-SA 4.0 license when loaded through the combined en configuration; see License.

from datasets import load_dataset

# 🌍 Language-level access (every corpus for that language)
ds = load_dataset("nvidia/Granary-v2", "de")

# 🎯 Corpus-specific access, named <lang>_<corpus>
ds = load_dataset("nvidia/Granary-v2", "de_mcv26")
ds = load_dataset("nvidia/Granary-v2", "en_librilight")
ds = load_dataset("nvidia/Granary-v2", "ja_mcv26")

# 📡 Streaming for large configs
ds = load_dataset("nvidia/Granary-v2", "en", streaming=True)

📊 Data Sample Structure

{
    "audio_filepath": str,            # Identifier used to locate source audio
    "duration": float,                # Duration in seconds
    "corpus": str,                    # source corpus name
    "source_lang": str,               # Source language code, e.g. "de"

    "tn_raw": str,                    # Primary transcript, text-normalized with P&C
    "itn_raw": str,                   # Inverse-text-normalized transcript
    "itn_clean": str,                 # ITN transcript after cleanup
    "captioning_text": str,           # Natural-language description of the audio
    "code_switched_text": str,        # Transcript with code-switching preserved
    "speech_qa_text": str,            # Q/A pair grounded in the audio ("SKIP" if N/A)

    "context_asr": dict | None,       # Coarse/fine context terms and entity categories
    "translations": dict,             # {"en": {"translation_raw": str,
                                      #          "translation_quality_score": float}}
    "preference_instructions": list,  # [{"prompt": str, "target": str, "tags": {}}, ...]
    "high_quality": bool | None,      # pipeline quality gate -- see note below
}

The pipeline uses 16 kHz mono audio for labeling. Source distributors can provide audio in other formats.

On high_quality. The flag records the pipeline's quality check. Approximately 165,300 hours in this public release pass it. Context and translation coverage vary by record. Early translation runs also included records outside the high-quality subset. Check each target for a usable value.

preference_instructions is a derived training view, not a core target: it packs the fields above into ready-to-train prompt/target pairs for instruction tuning. The pairs add no information beyond the other fields, so ignore them if you are training a plain ASR or AST model.

📄 License

This dataset is governed by the OpenMDW License Agreement, version 1.1, except for the People_Speech subset, which is governed by the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).

• OpenMDW-1.1: See LICENSE and the official agreement.
• People_Speech exception: See CC BY-SA 4.0 and the official license.

People_Speech refers to the released people_speech corpus, stored at en/people_speech/en_asr.jsonl.gz and available through configuration en_people_speech. CC BY-SA 4.0 also applies to these records when they are accessed through the combined en configuration. OpenMDW-1.1 applies to the other released manifests. See NOTICE for the subset mapping.

Obtain source audio separately from its distributors under the applicable upstream terms. The licenses above describe the dataset materials released in this repository.

📚 Citation

If you use Granary-v2, please cite the NVIDIA technical report below. The report describes the curation system and distinguishes the full processed collection from the public release subset.

@techreport{koluguri2026granaryv2,
  title       = {Granary-v2: Scalable Multilingual Speech Data Curation for Predictable and Controllable Model Outputs},
  author      = {Nithin Rao Koluguri and Meline Mkrtchyan and Nune Tadevosyan and Aaftab V and Sushmitha Deva and Fan Qian and Davit Karamyan and
                 Kunal Dhawan and Taejin Park and Piotr {\.Z}elasko and Arushi Dhingra and Viraj Karandikar and Shuoyang Ding and Slyne Deng and
                 Onur Yilmaz and Praateek Mahajan and Utkarsh Vaidya and Nikolay Karpov and Jagadeesh Balam and Boris Ginsburg},
  institution = {NVIDIA},
  type        = {Technical Report},
  year        = {2026},
  url         = {https://research.nvidia.com/labs/nemotron/files/granary-v2.pdf}
}

For background on Granary-v1, see the Granary-v1 paper in the References section.

Dataset Description:

Granary-v2 provides synthetic text labels and instruction pairs for multilingual speech. The repository contains compressed JSONL manifests. Obtain the source audio from the distributors listed in the Hours by corpus table. The Overview and data sample sections above describe the available targets and fields.

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

October 28, 2026

Version: 2.0

Previous Version(s): Granary-v1 (1.0).

Relationship to Previous Version(s): Granary-v2 replaces the labelling pipeline and expands the release to 48 languages and 16 source entries. It adds inverse text normalization, captioning, code-switching, speech-QA, context terms, and instruction pairs. Translations are stored in each record's translations field, keyed by target language, and run in both directions: non-English audio into English, and English audio into 31 target languages.

License/Terms of Use:

This dataset is governed by the OpenMDW License Agreement, version 1.1, except for the People_Speech subset, which is governed by the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).

See License, LICENSE, CC BY-SA 4.0, and NOTICE for the license texts and subset mapping. Obtain source audio separately from its distributors under their applicable terms.

Intended Usage:

Use the dataset for automatic speech recognition, speech translation in both directions (non-English audio into English, and English audio into 31 target languages), and speech instruction tuning. Additional targets support inverse text normalization, audio captioning, code-switching, speech question answering, and contextual speech recognition. Target availability varies by record; see the Data Sample Structure section above.

Dataset Characterization

• Data Collection Method: Manually-Collected. This release uses existing source corpora. The repository distributes manifests; obtain the audio from the source distributors.
• Labeling Method: Synthetic. The pipeline produces labels with automatic speech and text models.

Dataset Format

The repository contains compressed JSONL manifests (.jsonl.gz). Each line contains one utterance record. The manifest files for each language and corpus contain the available ASR and speech translation targets. Translations are stored in the translations field, keyed by target language.

See Repository Structure, Quick Start, and Data Sample Structure above for file paths, loading examples, and record fields.

Dataset Quantification

Measure	Value
Languages	48
Source entries	16
ASR hours	463,062
Hours marked high-quality by the pipeline	approximately 165,300
Speech translation hours (X→English)	92,496
Speech translation hours (English→X)	39,375

Speech translation uses audio already included in the ASR total. See Hours by corpus above for the corpus distribution.

Sources

The Hours by corpus table lists all 16 source entries, with links to their sources and their source-corpus licenses. This repository distributes manifests. Obtain the audio from each source distributor under that corpus's terms.

Reference(s):

• Nithin Rao Koluguri, Meline Mkrtchyan, Nune Tadevosyan, et al. (2026). Granary-v2: Scalable Multilingual Speech Data Curation for Predictable and Controllable Model Outputs. NVIDIA technical report.
• Granary-v1 dataset.
• Granary-v1 paper: Speech Recognition and Translation Dataset in 25 European Languages — describes the v1 pipeline, which this release supersedes.
• Granary-v2 curation pipeline in NeMo Curator.
• NVIDIA NeMo Toolkit — for training on these manifests.

The Citation section above provides the Granary-v2 NVIDIA technical report citation. The Granary-v1 paper describes the earlier release.

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.