Just yesterday, NVIDIA dropped something that makes most audio AI tools look like they're from a different decade. Nemotron-Labs-Audex-30B-A3B — or simply Audex — is a unified audio-text LLM that does it all: speech recognition, language translation, text-to-speech, and audio generation, all inside a single transformer decoder.
Why Audex Matters
Most audio AI systems are a Frankenstein of separate models — one for ASR, one for TTS, one for translation, one for music generation. Each one has its own architecture, its own latency, its own failure modes.
Audex smashes them together. Built on NVIDIA's Nemotron-Cascade-2-30B-A3B (a Mixture-of-Experts text backbone), it processes and generates audio natively — no glue code, no pipeline coordination.
The Architecture
- 30B total parameters, 3B active — MoE keeps inference fast and memory-light
- Single decoder handles both text and audio tokens in the same latent space
- Preserves the full text intelligence of Nemotron-Cascade-2 — no regression on language tasks
- A smaller 2B sibling is also available on Hugging Face for edge/on-device use
What It Can Do
- Speech recognition — multilingual, noisy environments, accents
- Speech-to-speech translation — hear your voice in another language, in real time
- Text-to-speech — natural prosody, controllable voices
- Audio generation — sound effects, music, environmental audio
Open Source, Open Weights
The model is live on Hugging Face under an open license. That means you can run it locally, fine-tune it, or embed it in your product without paying per-token API fees.
The Takeaway
Audex isn't just another model release — it's a signal that unified multimodal audio is becoming a commodity. NVIDIA took the MoE recipe that made Nemotron great at text and extended it into the audio domain without sacrificing either side. For developers building voice assistants, translation tools, or content creation pipelines, this is the kind of model that changes what's possible.
Check it out on Hugging Face: nvidia/Nemotron-Labs-Audex-30B-A3B
What are you building with Audex? Drop a comment below.
Top comments (0)