NVIDIA Unveils Groundbreaking AI Models and Dataset for Multilingual Speech Recognition

Aug 15, 2025
NVIDIA Blog
Article image for NVIDIA Unveils Groundbreaking AI Models and Dataset for Multilingual Speech Recognition

Summary

NVIDIA unveils Granary, a 1 million hour dataset spanning 25 languages, along with Canary-1b-v2 for multilingual speech recognition and translation, and Parakeet-tdt-0.6b-v3 for real-time transcription, pushing boundaries in AI language models.

Key Points

  • NVIDIA releases Granary, an open dataset containing around 1 million hours of audio for 25 European languages
  • NVIDIA unveils Canary-1b-v2, a high-accuracy model for multilingual speech recognition and translation, topping accuracy leaderboards
  • NVIDIA introduces Parakeet-tdt-0.6b-v3, a high-throughput model for real-time transcription of Granary's supported languages

Tags

Read Original Article