One checkpoint now does the job of three tools. On October 28, 2026, NVIDIA released Nemotron 3.5 Transcribe, an open-weight model that transcribes up to an hour of audio in 40 languages and tags each passage with the person speaking. Until now, a team building a voice product typically bolted together separate pieces for speech recognition, speaker identification and word alignment. NVIDIA has folded all of that into a single file.
The weights are free for commercial and non-commercial use under the OpenMDW-1.1 license. That matters: the shift here is from renting speech services to owning them.
What is inside
The model is a Mixture-of-Experts design with about 34 billion parameters in total, though only around 4.3 billion are active for any given step. Its name says 30B-A4B, which trips people up; the audio components and an extra prediction module account for the difference. A hybrid Mamba-Transformer backbone copes with the 65,536-token context that an hour of speech demands.
You steer it with a plain-text instruction. Ask for verbatim speech, a cleaned-up written version or a translation, and you get that. It can also return speaker labels and word-level timestamps. A frozen diarization component follows up to eight speakers at once, so the output shows who said what and when. NVIDIA trained it on roughly 1.72 million unique hours of audio.
How it scores
On the English Open ASR Leaderboard the model posts a mean word error rate of 4.71%. The next open model, Qwen3-ASR-1.7B, manages 4.94%. Across seven languages on the multilingual board, Nemotron lands at 4.62% against 6.49% for the same rival.
Commercial services still win. OtterAI speech-v3.1 leads English at 3.88%. NVIDIA grades each language as Leading, Production-ready or Supported, and puts Japanese in the middle tier. The weak spot is far-field Mandarin meetings: on the AliMeeting benchmark the model hit a 34.95% character error rate once speaker assignment is counted. These figures come from NVIDIA’s own reporting.
Hardware and limits
The main target is a single H100 with 80 GB of memory. Quantized FP8 and NVFP4 versions shrink the footprint to roughly 35 GB and 21 GB, and the NVFP4 build fits an RTX 5090. NVIDIA says a tuned setup on an RTX PRO 6000 Blackwell Server Edition card can chew through 2,238 seconds of audio every second. Those serving tweaks are not in the open release yet.
It is strictly an offline model. Feed it a recorded file, not a live call. Developers who want to adapt it to their own vocabulary can fine-tune it, and NVIDIA also released the Granary v2 dataset used in its development. The weights are on Hugging Face now, and a lighter quantized version is planned.
Sources
- nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B · Hugging Face
- NVIDIA has unveiled its transcription AI, 'Nemotron 3.5 Transcribe,' which can accurately identify speakers even in situations where many people are speaking simultaneously.
- RTX Spark Laptops Start at $2,599 as Nvidia Opens Speech AI – FourWeekMBA
- NVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters
- NVIDIA Nemotron 3.5 Transcribe: A Complete Guide to the Speech AI Model | HackerNoon
Photo: Zulfugar Karimov / Unsplash (Unsplash License)
