ReleasesSep 10, 2026 · 21:454 sources

Tencent Open Sources 1.5B AuK Speech Model Trained on 1.95 Million Hours

From huggingnews.com

Image: huggingnews.com
AI;DR

Tencent has open-sourced a 1.5 billion parameter AuK speech model trained on 1.95 million hours of data. The model uses a hybrid rectified-flow Transformer architecture to enable zero-shot voice cloning and audio cleanup from text prompts.

Read the sourceSign in to suggest

Sources

+ 3 more
  1. SUPPORT@chinazhidx“Trained on 1.95M hours of effective supervision across diverse audio data”x.com
  2. SUPPORT@papersdatacode“AuK-Flash runs in 4 steps without classifier-free guidance”x.com
  3. SUPPORT@genaispotlight“lightweight variant built for low-latency dubbing and game audio pipelines”x.com

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today