NVIDIA releases 2B audio-native model with open weights

NVIDIA releases Nemotron-Labs-Audex, a 2-billion-parameter audio-native model with open weights, according to a post on X. The post says the model handles transcription, translation, sound recognition, audio Q&A, text-to-speech and full speech-to-speech in one model, and that it "hears the world" rather than only words. A Hugging Face demo is linked for trying the model. The only source so far is this single X post, which provides no benchmark results.