NVIDIA's open-weight, promptable speech model turns recorded audio into text across 40 languages, translates 31 of them to and from English, labels up to eight speakers, and tops the open-weight tier of the Hugging Face Open ASR Leaderboard.
2026-10-28 · by Moe Ameen
On October 28, 2026, NVIDIA published Nemotron 3.5 Transcribe, an open-weight, promptable speech model that turns recorded audio into useful text. The released checkpoint, NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, is a mixture-of-experts model — roughly 30 billion parameters in total with about 4 billion active per token — that pairs a speech encoder with a hybrid Mamba-Transformer language model. Speaker tags come from NVIDIA's Sortformer and are produced during generation rather than as a separate pass.
On capability, it transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to about an hour and labels up to eight speakers, and it is built for recorded audio rather than live streaming. Beyond plain transcription it does speaker attribution, translation, word-level timestamps, and audio question answering (trained in English). "Promptable" means you steer the output with instructions — asking, for example, for a cleaned transcript or for specific terms to bias toward.
NVIDIA reports that the model leads the open-weight tier of the Hugging Face Open ASR Leaderboard, with roughly a 5.33% average word error rate on English at a high real-time factor, about 4.62% mean WER across a seven-language multilingual set, and 10.98% on long-form audio. It credits Granary v2 — an open multilingual speech dataset and curation pipeline of 463,062 hours across 48 languages, built with NVIDIA NeMo Curator — as part of the training recipe. Those accuracy figures are leaderboard- and vendor-reported; confirm them against the model card for your own audio.
The weights are free to download from Hugging Face and ModelScope, with hosted demos on Spaces and ModelScope Studios. The model serves on vLLM and SGLang-Omni from day one, hosted endpoints are available through Baseten, DeepInfra, and Simplismart, and it can be fine-tuned with NVIDIA NeMo Speech tools and ms-swift. NVIDIA did not publish standalone pricing, and a quantized build for faster local inference is listed as coming — confirm current details on NVIDIA's pages before depending on any figure.
The move this unlocks today is a cheap, multilingual front end to a content pipeline. Because the weights are open, you can transcribe a backlog of recordings — or a talk you gave in another language — without a per-minute API bill, and because the model translates 31 languages to and from English, a single foreign-language interview can come back as an English transcript you can actually work with. What it will not do is turn any of that into something you can post. That is where [Kompozy](/) picks up.
Take the speaker-labeled transcript, and the original long video if you have it, into Kompozy and the engine fans it into a week of finished assets across 18 formats — captioned [Clipped Shorts](/glossary/clipped-short) cut from the recording, a [Persona Short](/glossary/persona-shorts) delivering the key point, brand-exact [carousels](/glossary/hyperframes), quote graphics, a blog article, and an email newsletter — each rewritten under a [Persona Brief](/glossary/persona-brief) so a raw transcript becomes copy in your voice. [Autopilot](/glossary/autopilot) then schedules and publishes the set across the eight social platforms plus blog and email behind a per-post review gate. For a creator with a multilingual audience, the pairing is clean: transcribe and translate with Nemotron, then generate and ship with Kompozy.
It is an open-weight, promptable speech model NVIDIA released on October 28, 2026, that turns recorded audio into text. The checkpoint (30B-A4B) is a mixture-of-experts model that transcribes 40 languages, translates 31 to and from English, labels up to eight speakers, and also does word-level timestamps and English audio question answering.
It transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It is built for recorded files up to about an hour, not live streaming.
The weights are open and free to download from Hugging Face and ModelScope, and it serves on vLLM and SGLang-Omni. NVIDIA did not publish standalone pricing; check the model card for the exact license terms before building on it.
No. It produces transcripts and translations and nothing more — it makes no video, images, captioned clips, or scheduled posts. To turn a transcript into finished content across platforms you pair it with a generation-and-publishing engine like Kompozy.