Nemotron 3.5 Transcribe handles recordings, with task and language limits

NVIDIA [introduced Nemotron 3.5 Transcribe on October 28](https://blogs.nvidia.com/blog/local‑ai‑rtx‑spark‑laptops‑updates‑models‑agents‑games/) as an open‑weight model for working with recorded speech. A developer can ask for a transcript, a translation, speaker labels or an answer to a question about the audio. The useful distinction is the requested output: a readable transcript and a decision summary are different tasks, even when both start from the same meeting recording. The [model card’s prompt guide](https://huggingface.co/nvidia/NVIDIA‑Nemotron‑3.5‑Transcribe‑30B‑A4B) separates verbatim transcription, which keeps fillers and repetitions, from a readable style that removes fillers and accidental repetitions while retaining the spoken wording. It also describes supplying names and specialist vocabulary as context. That gives someone building a recording‑review workflow a way to specify the desired artifact rather than treating every request as a generic summary. These are NVIDIA’s documented capabilities; OpenTools has not run the model or evaluated its outputs.

The model card explicitly calls Transcribe an offline model, not a real‑time streaming speech‑recognition service. Its encoder uses 30‑second internal chunks, but NVIDIA says that is neither a 30‑second recording limit nor a streaming‑latency claim. The card lists support for recordings up to one hour and says the model is not guaranteed to generalize beyond its training limits.

For a team choosing between processing a saved interview and transcribing a live call as it happens, those are materially different requirements. Internal chunking does not turn this release into a live‑call service.

“Offline” also does not, by itself, mean that audio stays on a particular person’s laptop or inside an air‑gapped network. NVIDIA’s announcement describes deployment choices from desktop to cloud. Where audio goes therefore depends on the application and serving environment built around the weights, rather than the word offline alone.

The language matrix distinguishes transcription, translation to and from English, multispeaker transcription and code‑switching. Support is not identical across those columns. A language appearing in the matrix should not be read as a promise that every listed task works for it.

NVIDIA also says audio understanding, reasoning and following spoken instructions were evaluated only in English and may not work reliably in other languages. A product that accepts multilingual recordings for transcription cannot assume that answering questions about those recordings has the same evaluated language coverage.

Speaker‑attributed transcripts have a separate boundary. The card says speaker labels distinguish voices within a recording; they do not establish real‑world identity. Labels can organize who said what in an interview or meeting without proving that a voice belongs to a named person. That distinction should survive any downstream search, summary or review interface.

The [published quick start](https://huggingface.co/nvidia/NVIDIA‑Nemotron‑3.5‑Transcribe‑30B‑A4B) serves the checkpoint on an H100 with 80 GB of memory, using the specified vLLM and NeMo Speech setup. The card provides a separate hardware, runtime and operating‑system compatibility table. That example is a concrete deployment configuration, not evidence that any laptop or GPU can run the same checkpoint unchanged. The release is most clearly understood as a building block for saved‑recording workflows: choose the requested text output, check the relevant task and language coverage, then match the serving configuration to the actual hardware and data‑handling requirements. A transcript, a translated transcript and an answer generated from audio each need review appropriate to their intended use. Open weights provide deployment choice; they do not establish transcription accuracy, privacy controls or speaker identity on their own. NVIDIA’s [known‑risk notes](https://huggingface.co/nvidia/NVIDIA‑Nemotron‑3.5‑Transcribe‑30B‑A4B) warn that the model can misrepresent a recording while producing a plausible transcript or answer. They also say supplied context can pull the output toward incorrect or irrelevant terms. Vocabulary hints therefore remain inputs to review, rather than proof that a name or technical phrase was actually spoken. NVIDIA’s headquarters in Santa Clara, California, photographed August 4, 2018. Archive company context, not Nemotron 3.5 Transcribe’s interface or output, and not evidence of where recordings are processed. Photo: [Coolcaesar](https://commons.wikimedia.org/wiki/User:Coolcaesar); source perspective/crop adaptation by [Jacek Halicki](https://commons.wikimedia.org/wiki/User:Jacek_Halicki), [via Wikimedia Commons](https://commons.wikimedia.org/wiki/File:NVIDIA_Headquarters.jpg), [CC BY‑SA 4.0](https://creativecommons.org/licenses/by‑sa/4.0/). Cropped and downscaled for display; image adaptations remain under the same licence.