Nvidia releases open Nemotron 3.5 Transcribe for speech tasks

Nvidia has released Nemotron 3.5 Transcribe, an open speech recognition model that follows text instructions and combines transcription, translation and speaker recognition. The release also includes the Granary v2 dataset and tools for adapting the model to specialized fields such as medicine.

Capabilities and efficiency

Nvidia says the model transcribes 40 languages, translates 31 to or from English and identifies speakers in 13. It handles recordings up to an hour long with up to eight speakers. Simple instructions let the same recording become a transcript with speaker labels, an English translation or a summary. Technical terms such as drug names can be passed as hints, and the model is supposed to use them only if they appear. It can also answer questions about a recording's content.

The model has 30 billion parameters, but only about 4 billion are active per step, so it runs on a single Nvidia GPU. Nvidia says that covers hardware from the compact DGX Spark to data center systems. It processes 100 hours of audio in under three minutes on an RTX PRO 6000 Blackwell Server Edition at NVFP4 precision. It is built for recorded audio, does not support real-time processing, and its audio question-answering feature was trained in English.

Benchmark results

On Hugging Face's English Open ASR Leaderboard, Nemotron 3.5 Transcribe ranks first among open-weight models with an average word error rate of 5.33 percent. In Artificial Analysis' comparison, it scores 2.81 percent, close to Voxtral Small at 2.77 percent while using about a sixth of the active parameters. According to Nvidia, several widely used commercial speech services trail behind. It also edges out open models from Alibaba and IBM.

Its lead is largest in speaker recognition. VoiceArena tested 14 systems on conversations where speakers sometimes talk over each other, and Transcribe took first place. It dropped the fewest words, missing 8.32 percent compared with a median of about 31 percent. On David AI's benchmark across 21 languages, it posted the lowest error rate and finished just ahead of Nvidia's own dedicated model.

Dataset and adaptation

Alongside the model, Nvidia is releasing Granary v2, a dataset with about 844,000 hours of speech in 48 languages, up from roughly 643,000 hours in 25 languages. Each recording comes in verbatim and cleaned-up styles. Nvidia is sharing the tools used to build the dataset, and says the dataset does not fully match the model's training data.

On recordings of Indian doctors and medical students, a precise prompt cut the error rate from 30.8 to 23.8 percent, and fine-tuning reduced it to 11.8 percent. For speakers the model had not heard, Nvidia reports a drop from 22.44 to 12.17 percent. That put the adapted model ahead of an open medical speech model at 13.79 percent. The process took about 14 GPU hours on a single Nvidia GB300. Nvidia notes the specialization slightly raised the error rate on general English speech and recommends at least 100 real, verified recordings for testing custom versions.

Sources: The Decoder · TechCrunch · The New Stack