Skip to content

Dataset · nvidia

NVIDIA multilingual audio instruction dataset

nvidia/Granary-v2

Generated labels and instruction data for multilingual speech, covering ASR, translation, and speech question answering.

What's inside

The second release of NVIDIA's audio dataset featuring refined annotations across a wide range of source speech corpora. Along with transcriptions and translations, it includes labels for text normalization, audio captioning, code-switching, and speech QA pairs. Data is available in multiple languages including Arabic, Assamese, Bulgarian, Bengali, Czech, and Danish.

What you can do with it

  • Training and fine-tuning multilingual automatic speech recognition (ASR) and machine translation models
  • Training multimodal audio-text models for speech question answering (speech-QA)
  • Adapting models for code-switching and text normalization
Tasks
Automatic speech recognition, Translation, Audio text → text, Question answering
Languages
Arabic, Assamese, Bulgarian, Bangla, Czech, Danish, German, Greek and 4 more
Downloads this month
10.6K
Likes
61

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("nvidia/Granary-v2")