Nvidia releases all-in-one adaptable Nemotron 3.5 audio transcription and translation model
SiliconANGLE Kyt Dotson
Nvidia launched Nemotron 3.5 Transcribe, one model for transcription, translation and speaker labels. It handles 100 hours of audio in 40 languages on a single GPU.
Based on reporting by SiliconANGLE, Kyt Dotson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nvidia has put out Nemotron 3.5 Transcribe, an open-weight audio model that folds speech detection, transcription, speaker labeling and question answering into one pass. The pitch is simple: fewer moving parts, less glue code, and less time stitching together separate tools.
The model is a 30-billion-parameter mixture-of-experts system, but only 4 billion parameters are active at once. That lets it run on comparatively small hardware, while stronger machines can push it faster. Nvidia says it can work anywhere from the edge to the data center, depending on how it’s compressed.
The company says the model can handle up to 100 hours of audio, supports 40 languages and can label as many as eight speakers. It runs on a single Nvidia GPU, including Jetson AGX Thor, DGX Spark and Blackwell data-center GPUs. On a Blackwell Server Edition GPU with NVFP4 precision and one-quarter model compression, Nvidia says it can process 100 hours of audio in under three minutes.
Accuracy is part of the sales pitch too. Nvidia says the model posts a 30.8% word error rate out of the box, which can drop to 20.8% with careful prompting. On the Eka Care Indian medical speech benchmark, a fine-tuned version reached 11.8% word error rate. The model also takes prompt-based instructions such as transcribing and labeling speakers, translating audio into English, or answering questions about a meeting.
To support fine-tuning, Nvidia expanded its Granary dataset into Granary v2 using NeMo Curator. The public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1. Nvidia also says Granary v2 produces three transcript styles for each recording: fully verbatim, a version that turns numbers into symbols, and one that removes filler words. The model is available as a Hugging Face demo, with a checkpoint that developers can download and adapt.
My take — AI-written commentary, not fact-checked reporting
This is the sort of open model move that actually matters: not bigger for bragging rights, but more useful because it’s easier to bend to a real job. The industry loves pretending every audio product needs a whole stack of separate models; usually that just means more places for things to go wrong. One model doing the work, with prompts controlling the output, is far more practical than the usual AI fireworks.
Read more about this at: SiliconANGLE