NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, for processing recorded audio. Its model card describes transcription, speaker-attributed transcription, speech translation and audio question answering in one downloadable model. It is an offline model, so developers evaluating it for meeting archives or media processing should distinguish those uses from live transcription.
This article examines NVIDIA's supplied model documentation and a separate SALM adaptation study. No model run or independent reproduction of accuracy or speed results was performed for this article.
Check the language for each task
The model card lists transcription in 40 languages and recordings up to one hour. Translation and multispeaker transcription have their own language coverage. Support for transcribing a language does not establish support for every other task in that language.
Audio understanding, reasoning, and spoken instruction following were evaluated only in English. Speaker labels distinguish speakers within a recording; they do not establish real-world identity. A multilingual meeting application therefore needs to check its required language and task against the model card's support matrix.
The prompt guide lets developers request verbatim wording or text with fillers and accidental repetitions omitted. Text context can also supply vocabulary hints. An interview archive may need to preserve hesitations, while searchable meeting notes may benefit from omitting them. The prompt can state that choice.
Budget for the serving implementation
NVIDIA reports up to 2,238 seconds of audio processed per second on one RTX PRO 6000 Blackwell Server Edition GPU, using NVFP4 with DFlash2 speculative decoding. This figure uses the optimized vLLM implementation in Nemotron Transcribe NIM. These serving optimizations are not yet part of the open-source release.
The card separately reports a 0.13-second response for a 30-second recording with that optimized implementation. Neither figure establishes the response time for a single hour-long recording with the open implementation. NVIDIA describes the release-day NeMo Speech vLLM integration as intended for model development and expects it to be roughly two times slower.
The model card lists BF16, FP8 and NVFP4 variants. Its quick start serves NVIDIA-Nemotron-3.5-Transcribe-30B-A4B on one H100 80 GB with vLLM 0.32.0 and the latest NeMo Speech main branch. Check the precision, supported hardware and serving recipe together before using the published throughput to estimate deployment costs.
A separate adaptation example shows what to check
NVIDIA's medical transcription study offers an evaluation example for specialized vocabulary. The supplied study identifies its starting point as a SALM checkpoint but does not give an identifier tying it to Nemotron 3.5 Transcribe. Its results therefore do not establish how this release performs after adaptation.
The study split Indian medical speech recordings by speaker, using 855 clips for selection and reserving 2,764 for evaluation. On held-out speakers, it reports word error rate of 22.44% for the base checkpoint with its best prompt and 12.17% for the fine-tuned recipe.
Brand-name drug exact matches were 36.6% for the fine-tuned model and 38.3% for Parrotlet-a-en-5b. General-English word error rate on LibriSpeech test-clean also rose from 1.54% to 1.69%. These results illustrate why an overall transcription score needs a separate check of the names an application depends on.
For developers considering Nemotron 3.5 Transcribe, the next step is to evaluate real recordings in the intended language with the exact artifact and serving software they plan to deploy. Check transcription errors, important names and response time separately. The supplied documentation supports a recorded-audio evaluation; it does not supply an independently reproduced deployment budget.