NVIDIA releases Nemotron 3.5 Transcribe: 40 languages and 100 hours of audio in 3 minutes

NVIDIA has published Nemotron 3.5 Transcribe, an open speech-to-text model. It works with 40 languages, translates 31 of them, and can separate speakers in 13. Of roughly 34 billion parameters, about 4.3 billion activate per token, which makes the model fast: on a single RTX PRO 6000 it processed 100 hours of recordings in under three minutes. NVIDIA also released the Granary v2 dataset, which contains 463,000 hours of speech.

Russian gets the full set of tasks

Russian supports everything NVIDIA trained the model to do: transcription, translation, speaker separation, and mixed-language speech, where languages alternate within a single recording. That last capability matters for real-world calls, where Russian and English terms are mixed together and classic ASR systems often break down.

The model accepts files up to an hour long and adds word-level timestamps. You can steer it with text instructions, for example by asking it to transcribe a recording and label each speaker. There is also a mode for answering questions about the audio content, but only in English. The model does not process live streams; it is built for finished files.

Mamba, a transformer, and a mixture of experts

The architecture consists of a speech encoder and a hybrid language model in which Mamba layers alternate with transformer layers, and some blocks work as a mixture of experts. That explains the gap between 34 billion total parameters and 4.3 billion active ones. Memory has to hold all the weights, but the compute required per token is closer to that of a much smaller model.

The Artificial Analysis leaderboard shows this. Nemotron 3.5 Transcribe scores a 2.8% error rate there, and so does a competitor that needs roughly six times as many active parameters. At equal quality, the gain comes in inference cost.

The numbers and what ships

• Languages: 40 for transcription, 31 for translation, 13 for speaker separation.
• Recording length: up to one hour, with word-level timestamps.
• Parameters: 34 billion total, about 4.3 billion active per token.
• Composition: a speech encoder and a hybrid Mamba + transformer language model with a mixture of experts.
• Hugging Face open transcription leaderboard: first place among open models. The average word error rate was 5.33% on English, 4.62% on the multilingual portion, and 10.98% on long recordings.
• Artificial Analysis: 2.8% error rate.
• Speed: 100 hours of audio in under three minutes on a single RTX PRO 6000. That is more than 2000 times faster than real time.
• License: OpenMDW-1.1.
• Granary v2 dataset: 463,000 hours of speech, 48 languages, assembled from 16 open corpora.

The 10.98% figure on long recordings shows where room for improvement remains. An hour of audio with different voices and background noise still produces roughly twice as many errors as short segments.

The dataset deserves separate mention. NVIDIA published the descriptions and the processing pipeline, but the audio itself has to be downloaded separately, subject to the licenses of the source corpora. Even so, other teams get a ready-made recipe for building such a dataset, and they can reproduce it or adapt it to their own languages.

Who it's for and what it's good at

The model suits batch processing of archives: call recordings, podcasts, lectures, and interviews. Teams that produce subtitles or meeting minutes will find word-level timestamps and speaker labeling useful. Russian-speaking teams benefit from the mixed-speech support, and the open license allows deployment on their own servers, without sending recordings to an external API.

The limitations are also clear. The model is not suitable for voice assistants or live captions, because it has no streaming. Questions about audio work only in English. The weights, technical report, demo, and dataset are already on Hugging Face, so you can test quality on your own recordings right away, before building the model into a product.