NVIDIA's Audex Unifies Audio Understanding and Speech

NVIDIA has released Nemotron-Labs-Audex-30B-A3B, a mixture-of-experts model that folds audio understanding and speech generation into a single system. According to the model's Hugging Face page, it is designed as a unified audio-text architecture — able to both interpret incoming audio and produce speech, rather than splitting those jobs across separate specialized models.

The model uses a mixture-of-experts design with roughly 30 billion total parameters but only about 3 billion active per token, the arrangement its "30B-A3B" name signals. That approach keeps inference costs closer to a small dense model while giving the network a larger pool of specialized capacity to draw from, a pattern NVIDIA and others have leaned on across recent Nemotron releases.

Why it matters

Most production audio stacks still chain together distinct components — a speech recognizer, a language model, and a separate text-to-speech engine. A single model spanning both comprehension and generation could simplify those pipelines and reduce the latency and error accumulation that comes from stitching parts together.

• Unified audio-text handling for both understanding and speech synthesis
• Mixture-of-experts efficiency: ~30B total, ~3B active parameters
• Reasoning listed among its capabilities alongside audio and TTS

The release ships under a custom NVIDIA license rather than a standard open-source one, so teams will want to check the terms before building on it. As an initial version, Audex sets a baseline for a family that NVIDIA appears positioned to iterate on.