Nemotron 3.5 Transcribe: NVIDIA Open Speech Model, 40 Languages
Nemotron 3.5 Transcribe is NVIDIA's open-weight promptable speech model for transcription in 40 languages, translation in 31, and speaker labels in 13.
Early story. Some claims here are not officially confirmed yet. We update this post as it confirms.
- One checkpoint handles transcription, speaker labels, translation, and audio question answering via natural-language instructions.
- Hugging Face's Open ASR Leaderboard English track puts it first among open-weight models at about 5.33% average WER in the company blog's figure; Artificial Analysis places it on the open-weight accuracy–active-parameter frontier at roughly 4B active parameters of a ~30B MoE.
- Weights are on Hugging Face and ModelScope under OpenMDW-1.1, with Granary v2 data and NeMo Speech fine-tuning recipes released alongside.
in this block
Nemotron 3.5 Transcribe is NVIDIA's new open-weight speech model, and it tries to replace a stack of separate ASR, diarization, and translation tools with one promptable checkpoint. NVIDIA's research page and a Hugging Face blog dated October 28 say the model transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13, on recordings up to about an hour long.
What actually happened
NVIDIA published the model on October 28 as a speech-augmented language model. The research page describes a Transformer speech encoder paired with a frozen Nemotron 3 Diarization model and the Nemotron 3.5 Lightning 30B-A3B hybrid Mamba–Transformer mixture-of-experts LLM, for roughly 4.3B active parameters. The same checkpoint can do verbatim or cleaned transcripts, contextual ASR with entity lists, speaker-attributed output, speech translation, and spoken language ID, with style chosen by a text instruction per request.
The Hugging Face blog frames the pitch for developers: guide output with prompts such as "Transcribe this recording and label each speaker," "Translate this audio into English," or "What decisions were made in this meeting?" It says the model labels up to eight speakers, supports word-level timestamps, and is built for recorded audio rather than live streaming. Quantized versions for local RTX Spark, DGX Spark, Jetson AGX Thor, and Blackwell servers are described as coming soon; vLLM serving is available now for the released checkpoint.
TechCrunch's October 28 coverage placed the launch in a broader "voice AI" wave, noting NVIDIA's open transcription push. That is a different product line from the vision-language Nemotron we covered in Nemotron 3.5 Super VL, and separate from tabular work in NVIDIA Kumo.
Benchmarks and the open data angle
Company-reported leaderboard highlights, via the Hugging Face blog and NVIDIA page, include first place among open-weight models on the Hugging Face Open ASR English track, strong multilingual and long-form WER in their Open ASR comparisons, a tie for first among open-weight models on Artificial Analysis's AA-WER at 2.8% with about 4B active parameters, and top diarization results on David AI's 21-language track and VoiceArena's full-session English bench. NVIDIA also cites a medical fine-tuning case study where adaptation cut WER on held-out speakers by 45.8% relative to the best prompt.
Granary v2 is the open curation pipeline behind much of the training story. The public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1, with multiple transcript styles and task labels. Those figures describe the dataset release, not the full Transcribe training mixture.
Promptable ASR is useful only if you spot-check names, speaker turns, and answers against the audio.
Limits to keep in mind
The release is not a live-streaming engine in the launch materials. Language support differs by task, and audio question answering is trained in English. Leaderboard wins are impressive but still depend on the evaluation sets those boards use. Fine-tuning recipes on a single GPU are a real differentiator for domain jargon, yet they require labeled audio and engineering time.
Open weights under OpenMDW-1.1 lower the barrier to self-hosting, which matters for teams that cannot send calls to a hosted speech API. Throughput claims — such as processing large batches quickly on an RTX PRO 6000 Blackwell in NVIDIA's own tests — should be read as vendor benchmarks until you reproduce them on your hardware.
What to do as a reader
If you evaluate Nemotron 3.5 Transcribe, start from the official research page and the Hugging Face model card, not from social clips. Test your actual accents, overlap patterns, and vocabulary with the recommended prompts before you trust a production pipeline.
Compare open self-hosting cost against proprietary speech APIs on the same files. Keep a human review step for speaker attribution and proper nouns. Nothing here is investment advice, and speech models still invent confident wrong words.
Sources: NVIDIA Nemotron 3.5 Transcribe, Hugging Face blog and TechCrunch.
Not financial advice. DYOR, ser.
