Granary: 1 Million Hours of Speech for 25 European Languages

NVIDIA's Granary dataset provides approximately 1 million hours of high-quality speech data across 25 European languages for speech recognition and translation. Released in 2026, it consolidates multiple sources into a unified framework to support low-resource language modeling. The collection is designed for both Automatic Speech Recognition (ASR) and Automatic Speech Translation (AST) tasks.

Use Cases

• Training ASR models for low-resource languages using the 'mt' (Maltese) or 'lv' (Latvian) language labels
• Developing AST systems to translate spoken 'fr' (French) or 'de' (German) audio into text
• Fine-tuning multilingual models across the 25 European language categories provided in the unified framework

Strengths

• ~1 million hours of speech data
• Coverage of 25 European languages including low-resource variants
• Unified framework for ASR and AST tasks

Limitations

• Lack of detailed column-level metadata in the provided summary
• Potential variability in audio quality across consolidated sources

Provenance

Source

NVIDIA

Collection Method

Consolidation of multiple existing datasets under a unified framework

Freshness

Last updated March 2026.

Geography

Europe