Skip to content

Dataset · nvidia

NVIDIA clean multilingual speech metadata

nvidia/Granary-HiFi

Filtered metadata for a subset of VoxPopuli (10 to 100 million entries) for selecting high-quality audio in ASR.

What's inside

These are tabular and text metadata of Granary-v2 from NVIDIA, obtained after filtering the multilingual speech corpus VoxPopuli (including English, German, Czech, and other languages). There are no audio files inside—only annotated metadata that underwent sound event detection to filter out recordings with extraneous noise. The set size is 10 to 100 million rows.

What you can do with it

  • Filter the source VoxPopuli corpus for training ASR models on clean speech
  • Analyze speech coverage and noise levels in multilingual audio datasets
Tasks
Automatic speech recognition
Languages
Bulgarian, Czech, Danish, German, Greek, English, Spanish, Estonian and 4 more
Size
10M–100M rows
Downloads this month
2K
Likes
11

How to load it

from datasets import load_dataset

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