Audex-30B: NVIDIA's Neural Network for Audio and Speech

On July 7, 2026, NVIDIA released Audex-30B, a unified audio and speech system built on a Mixture of Experts architecture. It combines five types of audio tasks in one model without a significant loss of the base system's text abilities.

What Audex combines in one model

Before Audex, each audio task required its own specialized solution: one system for speech recognition, another for translation, a third for synthesis. NVIDIA has merged them into a single MoE architecture:

• Audio content understanding: analyzing the contents of sound files
• Automatic speech recognition (ASR)
• Speech translation into another language
• Speech synthesis: text-to-speech (TTS)
• Audio generation

The Mixture of Experts architecture lets the model activate only a portion of its parameters depending on the task at hand. Of the 30 billion total parameters, about 3 billion are active at inference, hence the suffix A3B in the full name. This approach reduces compute load at inference compared with a dense model of the same total size.

Why preserving text abilities matters

Adding a new modality to an already trained language model is traditionally accompanied by "catastrophic forgetting": the original text abilities degrade under the influence of the new fine-tuning. It is one of the key engineering challenges in building multimodal systems.

NVIDIA built Audex on top of Nemotron-Cascade-2, its own text base with strong language performance. The company states that the regression on text benchmarks was minimal: the audio capabilities were added without a noticeable drop in text quality.

The MoE architecture plays a key role here: the audio and text "experts" inside the model are partly isolated from each other, which reduces cross-interference between modalities during fine-tuning. As a result, the model stays competent in both text and sound, with no trade-off between the two areas.

Who this is relevant for

A single multitask model simplifies the development of voice and multimodal products. Instead of maintaining several specialized systems, a team works with one set of weights: it is easier to deploy, easier to update and easier to monitor. An update to the one model improves all five functions at once.

Potential use cases are broad: voice assistants that understand context, real-time transcription and translation systems, multilingual voice platforms for call centers, and audio content generation tools.

What this means

Audex continues the trend toward multimodal foundation models with full task coverage. Where "audio + text" used to mean stitching together several specialized systems, NVIDIA offers a single set of weights covering the whole range of audio tasks while keeping the text strength of the base model.

What can the Audex-30B neural network do?

It combines five types of audio tasks: audio content understanding, automatic speech recognition, synthesis, translation and an additional type of sound processing, all in one model that retains its text abilities.