Nvidia has released Granary v2, an open multilingual data-curation system and dataset manifest covering 463,062 hours of speech across 48 languages. Developed with NeMo Curator, the project addresses one of the most stubborn bottlenecks in enterprise voice tech: training models that can format, clean and translate audio based on plain instructions.

For enterprise IT and transformation leaders, automated speech recognition has historically forced an awkward trade-off. Off-the-shelf foundation models produce rigid transcripts that require downstream post-processing pipelines to remove disfluencies, reformat currency values, or add speaker attribution.

Granary v2 shifts that burden directly into the training stage by generating multiple aligned targets for the same audio segment.

Multi-Task Targets for Promptable Models

Unlike conventional datasets that treat a single reference text as the definitive ground truth, Granary v2 produces three distinct transcription conventions for every audio snippet:

  1. Spoken form (tn_raw): Preserves verbatim words with standard text normalization.
  2. Raw written form (itn_raw): Applies inverse text normalization while keeping conversational fillers.
  3. Clean written form (itn_clean): Strips disfluencies and formats numbers into standard digits and currency symbols.

In addition to transcription styles, the pipeline automatically extracts domain entity hints, cross-lingual translations (covering 142,800 hours into English and 55,000 hours of English into 31 languages), and speech-grounded question-answer pairs.

This data structure directly powers Nvidia’s Nemotron 3.5 Transcribe, allowing developers to steer output via natural language. A legal compliance team can prompt for word-for-word accuracy, while an executive assistant tool prompts for structured meeting notes, all using the same 34B checkpoint.

European Language Sovereignty and Data Governance

The expansion from 25 European languages in Granary v1 to 48 languages in v2 reflects broader enterprise pressure across Europe for sovereign, deployable open models. The public manifest includes all official EU languages alongside regional languages such as Catalan, Basque, and Galician.

From a regulatory standpoint, Granary v2 also provides clear audit trails. Under the EU AI Act and GDPR data governance requirements, enterprises deploying AI in high-stakes environments (such as healthcare consultations or HR interviews) face strict transparency rules regarding training data provenance.

Because Granary v2 releases its manifests, filtering criteria and NeMo Curator recipes under an open licence, European engineering teams can inspect the quality flags and run the exact same curation pipeline on their own private on-premises infrastructure.

Enterprise Reality: Batch Depth vs Real-Time Streaming

While Granary v2 marks a significant leap for speech engineering, organisations must account for its architectural boundaries:

  • Batch processing only: The pipeline and the models trained on it are optimised for recorded audio files up to one hour long, not sub-second live streaming. Real-time voice agents still require dedicated streaming engines.
  • Audio Q&A limitations: While transcription and speaker attribution span dozens of languages, audio question answering remains primarily trained on English corpora.
  • Hardware requirements: Running the full NeMo Curator pipeline across enterprise datasets requires Ray clusters and accelerated computing, making it a cloud or data-centre workflow rather than a desktop utility.

By releasing the recipes alongside the dataset manifest, Nvidia offers enterprise teams a clear framework: proprietary accuracy no longer requires closed APIs when open curation pipelines can match commercial benchmarks.