NVIDIA Nemotron 3.5 Transcribe brings promptable speech to one model
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, translation, and speaker attribution in one checkpoint. The model is available on Hugging Face and ModelScope. It transcribes 40 languages, translates 31 of them to or from English, and identifies speakers in 13 languages. NVIDIA says it handles recordings up to an hour long and labels up to eight speakers. TechCrunch reports that these tools target developers building apps for business meetings and medical transcription.
Developers control the model with natural language prompts. The same meeting recording can become a transcript with speaker labels, an English translation, or a summary of decisions made. You can also pass technical terms as hints, and NVIDIA says the model uses them only when the audio supports them.
In one line: NVIDIA released Nemotron 3.5 Transcribe, an open promptable speech model that does transcription, translation, and speaker attribution in one checkpoint, along with the Granary v2 dataset and fine-tuning tools.
What the model does
The model accepts recorded audio, not live streams. It supports word-level timestamps. The Hugging Face blog notes that audio question answering is trained in English. Language support differs by task, so some features may not be available for all 40 transcription languages. For speaker attribution, it works in 13 languages, according to NVIDIA.
A useful pattern from the model card: if you have domain jargon, you can prompt it before fine-tuning. For example, a healthcare prompt may list terms like HbA1c or metformin. The model is told to use these terms only when they appear in the audio and to preserve standard medical spelling.
Benchmarks and hardware
The model has 30 billion parameters, but only about 4 billion are active per token. NVIDIA says that lets it run on a single GPU, from DGX Spark to data center hardware. The Decoder reports that the model processes 100 hours of audio in under three minutes on an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision.
On the Hugging Face Open ASR Leaderboard for English, Nemotron 3.5 Transcribe ranks first among open-weight models with an average word error rate of 5.33 percent. For multilingual transcription, the Hugging Face blog lists a mean WER of 4.62 percent across seven languages. On Artificial Analysis' speech-to-text leaderboard, it scores 2.8 percent AA-WER, the benchmark's word error rate index. That is close to the best open model, Voxtral Small at 2.77 percent, but Transcribe uses about one sixth of the active parameters. For speaker attribution, it posts the lowest error rate on David AI's 21-language diarization leaderboard and ranks first in VoiceArena's 14-system comparison.
Fine-tuning with Granary v2
Alongside the model, NVIDIA released Granary v2, a dataset with about 844,000 hours of speech in 48 languages. The previous version had roughly 643,000 hours in 25 languages. Each recording has multiple transcription styles: verbatim with fillers and spelled-out numbers, and cleaned up with no fillers and digits. NVIDIA also shared the tools used to build the dataset, so developers can apply them to their own recordings.
NVIDIA tested fine-tuning on recordings of Indian doctors and medical students. A more precise prompt cut the error rate from 30.8 to 23.8 percent on the full benchmark. Fine-tuning brought it down to 11.8 percent. For held-out speakers, the relative improvement was 45.8 percent. That fine-tuned model beat an open medical speech model, which scored 13.79 percent. NVIDIA says the whole process took about 14 GPU hours on a single GB300. Some trade-offs appear: error rate rose slightly on general English speech, and the medical model stayed more accurate on brand-name drugs.
Limits to watch
This is a recorded audio model. It does not support real-time processing. NVIDIA recommends spot-checking names, speaker changes, and answers. Language support varies by task, and audio question answering is English only. The Decoder notes that the Granary v2 dataset does not fully match the model's training data. For custom versions, NVIDIA suggests setting aside at least 100 real, verified recordings for testing.
What I would try
For client delivery, recorded meeting notes and speaker labels are the first thing I would test. A model that can label speakers and accept a prompt with project terms could replace a couple of small scripts. I would start with a 30-minute call, pass a prompt that lists client-specific acronyms, and spot-check the output. Fine-tuning with Granary v2 looks useful only if we have enough labelled recordings. Without that, prompt-level control is the cheaper first step. Real-time voice agents are not in scope because this model works on recorded audio.