Nvidia combines transcription, translation, and voice in a single open-source model
Nemotron 3.5 Transcribe unites transcription, translation, and speaker identification in an open model. Nvidia accompanies it with the Granary v2 dataset and fine-tuning tools for specialized uses.
Nvidia has introduced Nemotron 3.5 Transcribe, an open voice model that transcribes, translates, and identifies speakers on recordings lasting up to one hour. According to the company, it handles 40 languages for transcription, translates 31 languages to or from English, and detects 13 languages for voice attribution. The launch also includes Granary v2, a massive audio dataset, and tools to adapt the model to specific sectors such as healthcare.
A model designed for meetings, not just demos
Nemotron 3.5 Transcribe does more than produce raw text. Depending on the instructions given, it can generate a transcription with speaker labels, an English translation, a summary of decisions made, or even answer questions about the content of a recording. The model also accepts technical vocabulary hints, for example drug names, with a simple rule: use them only if they actually appear in the audio.
On paper, the tool targets very concrete use cases: meeting minutes, transcription of conversations with overlapping voices, or processing specialized content where a simple ASR engine is not enough. ASR stands for automatic speech recognition. Nvidia pushes a simple idea here: a single model should be able to handle multiple audio tasks without requiring a chain of separate tools.
Large capacity, but streamlined execution
The model has 30 billion parameters, but only about 4 billion are active at each processing step. This architecture allows it, according to Nvidia, to run on a single GPU, from compact DGX Spark to data center hardware. On an RTX PRO 6000 Blackwell Server Edition, the company claims it processes 100 hours of audio in under three minutes at NVFP4 precision.
Benchmark figures also place Nemotron 3.5 Transcribe among the top performers. On Hugging Face’s English open-weight leaderboard, it ranks first with an average word error rate of 5.33%. On Artificial Analysis, it drops to 2.81%, very close to the best open model cited in the source, Voxtral Small, at 2.77%, while using about six times fewer active parameters. For speaker recognition, the margin seems even clearer: VoiceArena ranks it first on conversations with interruptions, with 8.32% missed words, versus a median close to 31%.
Granary v2, the working base for specialized uses
Along with the model, Nvidia releases Granary v2, a dataset gathering about 844,000 hours of speech in 48 languages. The previous version contained about 643,000 hours in 25 languages. Each recording is offered in several transcription styles: one faithful word-for-word, another cleaned of hesitations and converting numbers into digits. The company also shares the corpus-building tools so developers can reapply them to their own data.
The benefit is clear for teams wanting to adapt the model to a narrow domain. Nvidia says it tested fine-tuning on recordings of doctors and medical students in India. Across the benchmark, a better prompt reduces the error rate from 30.8% to 23.8%, then fine-tuning brings it down to 11.8%. For speakers unknown to the model, the error drops from 22.44% to 12.17%, a relative improvement of 45.8%. The procedure reportedly took about 14 GPU hours on a Nvidia GB300.
The classic downside of specialized models
This adaptation comes at a cost. Nvidia indicates that the error rate slightly rises on general English, while the medical model remains more reliable on brand-name drug names. The company also recommends keeping at least 100 real and verified recordings to validate a customized version. In other words: benchmarks shine quickly, but real life still has a knack for complicating things.
The system has other limits. Some functions vary by language, and the audio question-answering was trained in English. The model is designed for recorded audio, not real-time. Nvidia therefore advises manually checking proper names, speaker changes, and generated answers. Nemotron 3.5 Transcribe follows Nemotron 3 Diarization, introduced in September and dedicated to voice identification but without integrated transcription.
Key points
- 40 languages transcribed, according to Nvidia.
- 844,000 hours of audio in Granary v2.
- 5.33% average error on Hugging Face.
- 8.32% missed words on VoiceArena.
- 14 GPU hours for medical fine-tuning.
By the numbers
- 30 billion parameters in total, with 4 billion active per step.
- 31 languages translated to or from English.
- 48 languages covered by Granary v2.
- 100 hours of audio processed in under three minutes, according to Nvidia.