Zhifeng Kong, Sang-gil Lee, Jaehyeon Kim, Boxin Wang, Zihan Liu, Sungwon Kim, Yang Chen, Arushi Goel, Rajarshi Roy, Wenliang Dai, Zhuolin Yang, Yangyi Chen, Dongfu Jiang, Sreyan Ghosh, Tuomas Rintamaki, Andrew Tao, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping
5 min
As large language models (LLMs) evolve toward artificial general intelligence (AGI), integrating multimodal capabilities—specifically audio understanding and generation—without degrading the model's core text-based reasoning and knowledge capabilities remains a significant challenge. This paper introduces Nemotron-Labs-Audex-30B-A3B (Audex), a unified model designed to handle audio and text inputs and outputs within a single Transformer decoder architecture.
Audex is built upon the Nemotron-Cascade-2-30B-A3B, a text-only Mixture-of-Experts (MoE) LLM. The authors employ a unified design where audio inputs are processed by an audio encoder and projected into the text embedding space, while text and quantized audio tokens are treated uniformly during autoregressive generation. The training process involves a multi-stage supervised fine-tuning (SFT) curriculum, followed by text-only Cascade Reinforcement Learning (RL) and multi-domain on-policy distillation. This approach ensures that the model learns new multimodal tasks while maintaining its original text intelligence.
Audex achieves state-of-the-art results in general audio understanding, speech recognition, speech translation, text-to-speech, and audio generation. Crucially, it maintains performance parity with its text-only backbone on complex reasoning benchmarks (e.g., AIME, HMMT, and LiveCodeBench), avoiding the performance degradation often observed in other multimodal models. The authors demonstrate that a multi-stage SFT strategy is superior to a consolidated single-stage approach, as the latter can disrupt the model's long-context attention mechanisms.
This work provides a scalable, open-research-friendly architecture for building native multimodal LLMs. By proving that audio intelligence can be integrated into a high-performing text LLM without sacrificing reasoning or agentic capabilities, the authors offer a blueprint for future AGI development that treats audio as a first-class modality alongside text.
Audio intelligence involves understanding, reasoning about, and generating both audio and speech. In this work, we introduce Nemotron-Labs-Audex-30B-A3B (Audex), a unified audio-text LLM built on Nemotron-Cascade-2-30B-A3B, a strong text-only MoE LLM. Audex adopts a simple unified design with a single Transformer decoder: audio inputs are encoded and projected into the text embedding space, while text tokens and quantized audio output tokens are treated uniformly during generation. This architecture enables strong audio-text fusion, seamless multimodal generation, and compatibility with standard LLM training and inference infrastructure. For training, we meticulously curate audio-text datasets comprising 157.4B audio tokens and 320.5B text tokens. We apply multi-stage supervised training on these datasets, followed by text-only Cascade RL and multi-domain on-policy distillation. Audex delivers state-of-the-art audio understanding, speech recognition and translation, text-to-speech, audio generation, and speech-to-speech generation, while preserving very compelling reasoning, alignment, knowledge, long-context, and agentic capabilities of its text-only LLM backbone with marginal or no regression. We release the model checkpoints to facilitate open research.
Alex: And what happens at the very end of training?
Sam: They finish with a round of reinforcement learning aimed specifically at text reasoning. Think of it as a final exam in logic, taken after all the audio training is done. By sharpening the model's reasoning skills at the end, they effectively re-anchor its intelligence—making sure the audio work hasn't quietly eroded what was there before.
Alex: And does that actually hold up? Does the model stay sharp at reasoning?
Sam: That is the most notable finding the paper reports. On complex reasoning benchmarks, the audio-capable version performs almost identically to the text-only version. The authors argue this supports their central claim: if you treat audio as a language rather than a foreign add-on, you don't have to compromise the model's core intelligence to make it listen.
Alex: What can it actually do, in practical terms?
Sam: The paper describes a model that can transcribe speech, answer questions about audio content, generate sound effects, and work through complex coding or logic problems—all within the same system. The point isn't that any one of those tasks is new. Other models can do individual pieces. The claim is that this one does all of them without the usual trade-off.
Alex: It's a bit like the difference between a Swiss Army knife where every tool works properly, versus one where adding the scissors means the blade gets duller.
Sam: That's a fair way to put it. And the paper is careful not to overstate things. The authors acknowledge this is an early step. The approach works within the constraints they tested, but there's no suggestion it solves every challenge in multimodal AI. It's a meaningful result, not a final answer.
Alex: So the contribution here is really about the method—showing that the way you integrate audio matters enormously, and that a unified approach can preserve what makes a model useful in the first place.
Sam: Exactly. The architecture, the training sequence, the final reasoning refinement—each piece is designed to solve the same underlying problem: how do you add a new sense without losing your mind? Based on what this paper reports, treating sound as language rather than as something separate is a promising answer to that question.
Alex: Thanks for listening to ResearchPod.