NVIDIA's Nemotron 3.5 Transcribe Brings Promptable ASR

NVIDIA has released Nemotron 3.5 Transcribe 30B-A4B, a speech recognition model that leans on a mixture-of-experts design to keep inference light while scaling capacity. The model totals roughly 30 billion parameters but activates only about 4 billion per token, a now-familiar pattern that aims to deliver larger-model quality at a smaller compute cost.

The headline feature is that it is promptable: rather than a fixed transcription pipeline, the model can be steered toward different tasks through instructions. According to NVIDIA's release write-up, that includes multilingual transcription, speech translation, and speaker diarization from a single system.

What it does

• Multilingual automatic speech recognition
• Speech-to-text translation across languages
• Speaker diarization to separate who said what
• Prompt-based control over the output task

Combining these capabilities in one promptable model is the interesting part. Teams typically stitch together separate systems for transcription, translation, and diarization; folding them into a single checkpoint simplifies deployment and could reduce the glue code and latency that usually come with multi-stage audio pipelines.

Why it matters: open ASR has lagged behind the pace of open text and image models, and a sparse, instruction-driven option from a major vendor gives developers a flexible building block for captioning, meeting tools, and voice interfaces. The model ships under a custom NVIDIA license, so prospective users should review the terms before building on it.