Image: NVIDIA
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that turns recorded audio into text and can also label speakers, translate and answer questions about a recording. The company announced the model on X on October 28 alongside a detailed blog post on Hugging Face.
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
— NVIDIA AI (@NVIDIAAI) October 28, 2026
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
What it does
Transcribe is a mixture-of-experts model with 30 billion parameters, of which about 4 billion are active at a time, according to NVIDIA.
NVIDIA says the model transcribes 40 languages, translates 31 of them to or from English and attributes speakers in 13. It handles recordings up to an hour long and can label up to eight speakers, including when people talk over each other.
The model is promptable, so users steer it with plain instructions. NVIDIA's examples include asking it to label each speaker, translate a recording into English or list the decisions made in a meeting. In a demo video, the company shows it translating a meeting transcript into Spanish and then Japanese and redacting an email address as personal data. Prompts can also supply expected terms, such as drug names in a medical discussion.
How it scores
NVIDIA says Transcribe ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English, with a 5.33% average word error rate. On Artificial Analysis, the company reports a 2.81% word error rate, close to the 2.77% of Mistral's Voxtral Small, which uses about 24 billion active parameters.
Image: NVIDIA
NVIDIA also says the model ranks first in VoiceArena's 14-system comparison for speaker attribution and posts the lowest error rate on David AI's multilingual diarization leaderboard. These are company-reported results.
Who it is for and where it runs
The model targets developers building apps for meetings, calls and medical dictation, as TechCrunch noted. NVIDIA says it runs on a single NVIDIA GPU, from Jetson AGX Thor and DGX Spark up to Blackwell data center GPUs. On one RTX PRO 6000 Blackwell Server Edition GPU, the company says it processes 100 hours of audio in under three minutes, measured on short clips at high concurrency. Developers can serve it with vLLM or run it locally with NeMo-Speech.cpp.
For specialist vocabulary, NVIDIA released Granary v2, an open speech dataset, along with NeMo Speech fine-tuning tools. In a medical speech case study, NVIDIA says fine-tuning cut the word error rate by 45.8% relative to the best prompt.
Availability and limits
The weights are available now on Hugging Face and ModelScope, and a live demo lets anyone try a recording. The model card lists the OpenMDW 1.1 license and says the model is ready for commercial or non-commercial use. NVIDIA says the weights, data and recipes are open.
There are caveats. Transcribe is built for recorded audio, not live streaming. Language support differs by task, and audio question answering is trained in English. NVIDIA itself advises checking names, speaker changes and answers against the original recording. SiliconANGLE also covered the launch. The leaderboard positions above reflect rankings at the time of this writing.
Comments