NVIDIA Nemotron 3.5 Transcribe: open 30B model handles 40 languages, diarization and translation

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight, promptable speech model that combines transcription, speaker attribution (diarization), translation and audio question answering in a single checkpoint. It is available on Hugging Face and ModelScope. NVIDIA published the announcement on October 28, 2026.

Capabilities

The model 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. Word-level timestamps are supported. Output is steered with natural-language instructions, for example asking for speaker labels, an English translation, or a summary of decisions made in a meeting. Users can also supply domain terms in the prompt, such as medical vocabulary, and specify how numbers should be written.

Language support differs by task, and audio question answering is trained in English. The model card includes recommended presets, a prompt guide and a task-by-language matrix. NVIDIA advises spot-checking names, speaker changes and answers against the recording.

Architecture

A speech encoder feeds a hybrid Mamba-Transformer mixture-of-experts language model. About 4B of the model's 30B parameters are active per token. Sortformer speaker cues reach the language model alongside the acoustic features, so the model tags each turn during generation rather than relying only on a later step to attach speakers to words.

Benchmark results (reported by NVIDIA)

• Hugging Face Open ASR Leaderboard: first among open-weight models for English at 5.33% average WER (API models excluded). Multilingual mean WER is 4.62% across seven languages and long-form mean WER is 10.98%, the lowest among open-weight models in the model card's comparison.
• Artificial Analysis Speech-to-Text: ties for first among open-weight models at 2.8% AA-WER, with about 4B active parameters versus 24B for the model it ties. On AA-AgentTalk, which makes up 50% of the AA-WER index, it scores 1.8% WER versus 2.3% for the next-best open-weight model. NVIDIA says it ranks ahead of several widely used proprietary speech APIs on Artificial Analysis's all-model chart.
• David AI DAI-ASR-I18N: lowest diarization error rate of 14 systems across 21 languages at 18.26%, ahead of NVIDIA's dedicated Nemotron 3 Diarization at 18.38%. On the combined transcription-and-speaker-attribution track it ranks first among evaluated open-weight models and second overall.
• VoiceArena Diarization-Bench: lowest cpWER of 14 systems at 32.4%, ahead of NVIDIA Multitalker Parakeet with Nemotron 3 Diarization at 34.2%. Deletions are 8.32% of reference words, versus a median of about 31%.

NVIDIA also cites Ocular AI's Converse-STT evaluation, which ranks the model the top open model, and Josh Talks' Voice of India leaderboard for Indian-language results.

Granary v2 and fine-tuning

NVIDIA is also releasing Granary v2 multilingual data on Hugging Face and ModelScope, plus a reusable curation pipeline. Together with NeMo Speech fine-tuning tools, these let developers define the written form they need and adapt the model to their own speech. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt.

Deployment

The model runs on a single NVIDIA GPU, from Jetson AGX Thor and DGX Spark to Blackwell data-center GPUs. Agora, a real-time voice and video platform, used the word-level timestamps in a classroom demo that shows a transcript beside the replay of a recorded class.

What this means

The release puts several speech tasks behind one prompt interface: transcription, speaker labels, translation and audio Q&A. Because only about 4B of 30B parameters are active per token, the model's accuracy-to-compute ratio is the basis of NVIDIA's single-GPU claim and of its position on Artificial Analysis's open-weight frontier. The medical case study shows the intended workflow: use prompts and term lists first, then fine-tune with Granary v2 and NeMo Speech when prompting alone is not enough, which NVIDIA reports reduced WER by 45.8% relative to the best prompt in that case. Teams should note the task-by-language limits, notably that audio Q&A is trained in English, and verify results on their own audio, as NVIDIA itself recommends.