Granary-v2: what is in NVIDIA’s public speech-label release

NVIDIA’s Granary‑v2 release supplies generated labels and instruction–target pairs for multilingual speech training. Published on October 28, the project overview describes a full processed collection of approximately 843,600 unique audio hours. The public dataset card describes labels for 463,062 hours of source speech across 48 languages.

Those figures describe different scopes. The larger figure belongs to the collection processed by the curation system; it is not the amount of audio bundled into the public download. The public release is a subset of generated labels, with multiple representations and tasks attached to recordings. For a developer, the immediate question is which released records and targets are usable for the intended training job.

The Hugging Face repository contains compressed JSONL label manifests, not audio files. Each record’s audio_filepath is an identifier used to locate a recording from its source corpus. Developers obtain the corresponding audio separately from the distributors identified in the dataset card, under those distributors’ access requirements and terms.

The released labels and instruction pairs have their own stated terms: OpenMDW 1.1, except for the People’s Speech subset, which the card places under CC BY‑SA 4.0. That exception also applies when those records are accessed through the combined English configuration. The card separates these released‑material terms from the licenses and access rules for the source audio. This report describes that distinction; it does not independently review the full license agreements or establish a particular project’s compliance.

The practical preparation has two parts: identify the manifest and its applicable released‑material terms, then establish how the corresponding audio can be obtained and used. A successful manifest download settles neither the audio acquisition nor every reuse condition.

Granary‑v2 can attach normalized and inverse‑normalized transcripts, captions, code‑switched text, speech questions and translations to a recording. Coverage varies by record. The card’s schema and coverage notes identify several states a training‑data loader needs to preserve.

Field	State described in the card	What to check before using it
translations / translation_raw	translation_raw under a target‑language key can be an empty string after a candidate is rejected	Check translation_raw for a nonempty value, rather than accepting the key alone
speech_qa_text	SKIP means a speech question–answer target is not applicable	Keep the sentinel separate from a usable question and answer
context_asr	Context terms can be absent; the schema permits null	Preserve missing context instead of inventing it
high_quality	A nullable flag records the pipeline’s own quality gate	Treat it as a pipeline‑filter result, not human verification

The card says approximately 165,300 hours in the public release pass its quality gate. That is another subset definition, rather than an independent finding that every target in those records is complete. It specifically advises checking each target for a usable value.

Translation targets also reuse audio already counted in the speech‑recognition total. A recording translated into multiple languages does not become multiple new recordings or extra source‑audio hours. That matters when estimating the material available for one task or reporting the release’s size.

The card describes `preference_instructions` as a derived training view: it packages existing fields into prompt–target pairs without adding new information. A developer training a plain speech‑recognition or speech‑translation model need not treat those instruction pairs as an additional underlying dataset. [Source: NVIDIA dataset card](https://huggingface.co/datasets/nvidia/Granary‑v2) NVIDIA’s [project page](https://research.nvidia.com/labs/nemotron/Granary‑v2/) links the curation pipeline built with NeMo Curator and model‑training recipes. The dataset card says the pipeline source is public, while full end‑to‑end runs and per‑stage scripts will follow. Those are different stages of tooling availability. No code or audio was downloaded or executed for this report. A useful first pass therefore checks the released manifest’s population, the audio acquisition route, the specific target’s missing‑value rules and the chosen training view. This is a report on what the current release contains and how its documented fields differ, rather than a benchmark claim about the models trained on it. AI‑generated conceptual illustration of one recording with several text‑label views and an empty target. The reel, cards and marks are symbolic, not actual Granary‑v2 audio, transcripts, records or software. Original artwork by OpenTools Team, based on the [Granary‑v2 release](https://research.nvidia.com/labs/nemotron/Granary‑v2/).