NVIDIA announced Nemotron 3.5 Transcribe, an open-weight speech model, in a post published October 28, 2026 on Hugging Face. According to the announcement, the model combines transcription, speaker attribution (diarization), translation and audio question answering in a single checkpoint. NVIDIA states it 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 labels up to eight speakers. Developers can guide output using natural-language instructions and adapt recognition to domain vocabulary and accents. NVIDIA reports the model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English at 5.33% average word error rate (WER), and ties for first on Artificial Analysis's leaderboard at 2.8% AA-WER with about 4B active parameters. The company says only about 4B of the model's 30B parameters are active per token via a hybrid Mamba–Transformer mixture-of-experts design. NVIDIA also released Granary v2, a multilingual dataset and curation pipeline, plus NeMo fine-tuning tools. In a stated medical case study, fine-tuning cut WER on held-out speakers by 45.8% relative to the best prompt. NVIDIA notes users should spot-check names, speaker changes and answers against recordings. Benchmark claims are reported by NVIDIA and Artificial Analysis.
- Open-weight model handles transcription, diarization, translation and audio Q&A; 40 languages, 31 translation pairs, 13 for speaker attribution
- Handles recordings up to one hour and labels up to eight speakers
- NVIDIA reports 5.33% average WER (first among open-weight on Hugging Face Open ASR, English) and 2.8% AA-WER on Artificial Analysis
- Granary v2 dataset and NeMo fine-tuning tools released; medical case study cut WER 45.8% relative on held-out speakers
What it means for you
If your organization transcribes audio — calls, meetings, dictation, classes — this is a capable open model you can run yourself and adapt to your own vocabulary, rather than renting a black-box service. The practical draw is control: you choose where the audio is processed and can prompt it to spell your industry's terms correctly. It is not plug-and-play, though — using it well requires technical staff to set up and fine-tune, and NVIDIA itself says every transcript needs human spot-checking of names and speaker labels.
Try this
If transcription matters to your work, write down what 'good' actually looks like for you — do you need speaker labels, correct spelling of specific terms, timestamps? That one-page spec tells you whether your current tool falls short before anyone evaluates a new model.
Who should care
Teams that process sensitive or jargon-heavy audio where sending recordings to a cloud service is a problem — think healthcare dictation, legal, or internal meetings — and that have technical people able to deploy and tune an open model.
Skip this if
Your existing transcription tool already handles your audio acceptably, or you have no one in-house to run and maintain a self-hosted model. For most teams, the off-the-shelf service you already use is fine.
Sources: Hugging Face — read the original