NVIDIA's open-weight, promptable speech model that transcribes 40 languages, translates 31 to and from English, and labels up to eight speakers from recorded audio.
Last verified · 2026-10-28 · by Moe Ameen
NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable speech model released October 28, 2026. 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 as the model generates, not as a separate step.
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 translation, word-level timestamps, and English audio question answering, and being "promptable" means you can steer the output with instructions.
NVIDIA reports that it leads the open-weight tier of the Hugging Face Open ASR Leaderboard — roughly a 5.33% average word error rate on English and about 4.62% across a seven-language multilingual set — with a training recipe that draws on the open Granary v2 speech dataset. Those are leaderboard- and vendor-reported figures; confirm them on the model card for your audio.
You run it as a model, not an app. The weights are free to download from Hugging Face and ModelScope, it serves on vLLM and SGLang-Omni, hosted endpoints are available through Baseten, DeepInfra, and Simplismart, and it fine-tunes with NeMo Speech tools and ms-swift. For a creator, its value sits at the top of the funnel: it is a fast, cheap-to-run way to get accurate, speaker-labeled, multilingual transcripts out of recordings. It writes no posts and renders no media — it is an input tool, and a strong one.
Nemotron 3.5 Transcribe is at its best on the hard audio most transcription tools punt on: a multi-guest panel and a recording in a language that is not English. It labels up to eight speakers and translates 31 languages to and from English, so a two-hour panel comes back segmented by who said what, and a foreign-language interview comes back as an English transcript you can read. Neither is a post yet — that is where [Kompozy](/) turns the recording into a week of published content.
Because the transcript knows where each speaker's best lines fall, Kompozy's own clipping can cut a [Clipped Short](/glossary/clipped-short) per guest and burn in branded captions, using the timestamps as the map of the good moments. Feed the same transcript in as a source and Kompozy drafts under your [Persona Brief](/glossary/persona-brief) so a panel discussion becomes on-brand [carousels](/glossary/hyperframes), quote graphics, a blog recap, and a newsletter — and when the source was translated, you generate and publish in the language your audience actually reads. [Autopilot](/glossary/autopilot) reframes each output to 9:16, 1:1, and 16:9 and ships it across the eight social platforms plus blog and email behind a per-post review gate. Nemotron untangles the speakers and the language; Kompozy makes the media and distributes it.
NVIDIA's open-weight, promptable speech model, released October 28, 2026. The 30B-A4B mixture-of-experts checkpoint turns recorded audio into text across 40 languages, translates 31 to and from English, labels up to eight speakers, and also does word-level timestamps and English audio question answering.
NVIDIA reports roughly a 5.33% average word error rate on English and about 4.62% across a seven-language multilingual set, leading the open-weight tier of the Hugging Face Open ASR Leaderboard. Those are leaderboard- and vendor-reported figures, so real-world accuracy varies with audio quality — treat them as a benchmark.
The weights are free to download from Hugging Face and ModelScope, and it serves on vLLM and SGLang-Omni, so you can self-host it on a GPU. If you would rather not run it yourself, hosted endpoints are available through Baseten, DeepInfra, and Simplismart. Confirm current details and the license on the model card.
No. It produces transcripts and translations and nothing more — it renders no video or images and publishes nothing. To turn a transcript into finished posts across platforms, pair it with a content engine like Kompozy.
It is a strong first step, especially for panels and foreign-language audio: speaker labels and translation make a multi-guest or non-English recording easy to mine for quotes and clips. The slicing, captioning, and publishing still happen in a separate tool such as Kompozy.