NVIDIA opens Nemotron 3.5 Transcribe weights: one model for transcription, diarization and audio Q&A

NVIDIA released Nemotron 3.5 Transcribe as an open-weight speech model that handles transcription, speaker attribution, translation, and audio question answering in one checkpoint. The model transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13, processing recordings up to an hour long with up to eight speakers. NVIDIA published the weights as nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B on Hugging Face and ModelScope, alongside the Granary v2 data pipeline and fine-tuning recipes that let developers adapt it to their own vocabulary.

Baseten lists the model in its production library under the same identifier, an independent signal that the checkpoint is real and deployable now — not a preview. The vendor blog says vLLM can serve it today, and NeMo-Speech.cpp runs it locally.

Four speech tasks, one promptable checkpoint

Instead of wiring separate ASR, diarization, and translation components together, you give the model a natural-language instruction with the audio: "Transcribe this recording and label each speaker," "Translate this audio into English," or "What decisions were made in this meeting?" Domain jargon can be injected as context terms in the prompt — the healthcare example in the post feeds it terms like HbA1c and metformin with instructions to use them only when the audio supports them. Word-level timestamps are supported, which is what the live-class platform Agora highlights in a demo turning recorded classes into searchable transcripts.

Two constraints matter before you build on it. Transcribe is built for recorded audio, not live streaming. And support is uneven: the model card's task-by-language matrix shows different coverage per task, and audio question answering is trained in English.

What the leaderboards show

In its official announcement, NVIDIA reports first place among open-weight models on the Hugging Face Open ASR Leaderboard's English track at 5.33% average WER, 4.62% mean WER across seven languages in the multilingual track, and 10.98% on long-form — the lowest in the model card's comparison.

On Artificial Analysis, which independently ranks speech-to-text models by its own AA-WER index, the model reportedly ties for first among open weights at 2.8%, with roughly 4B active parameters versus 24B for the model it ties: a hybrid Mamba–Transformer mixture-of-experts language model behind the speech encoder.

NVIDIA's blog also claims the lowest diarization error rate (18.26% DER across 21 languages) on David AI's DAI-ASR-I18N leaderboard and the best cpWER (32.4%) in VoiceArena's 14-system Diarization-Bench. The specific figures are vendor-reported; the model itself is verifiable on the third-party platforms above.

For hardware, NVIDIA says one RTX PRO 6000 Blackwell Server GPU with NVFP4 processed 100 hours of audio in under three minutes at high concurrency, and the model runs from Jetson AGX Thor and DGX Spark up to data-center GPUs.

Granary v2: the open data behind it

Granary v2, the dataset and curation pipeline on Hugging Face, grew from 25 languages in v1 to a public manifest covering 463,062 hours across 48 languages and 16 source corpora. Each recording carries three written forms: verbatim with disfluencies, and normalized spelling with and without them. Labels include captions, Q&A pairs, biasing terms, and quality-scored translations. NVIDIA notes the manifest describes the dataset release, not Transcribe's full training mixture.

The pipeline is public too — built on NeMo Curator, runnable with Ray and NVIDIA Dynamo — while source audio must be downloaded separately under each corpus's original license. A HiFi filter drops noisy or overlapping segments, shrinking VoxPopuli from 354k to 111k hours.

Fine-tuning cut medical dictation errors by 45.8%

The medical case study is the practical payoff of the open pipeline. On Eka Care's open benchmark of Indian medical speech, prompting alone cut WER from 30.8% to 23.8%; adding domain fine-tuning brought it to 11.8% on the full set. On held-out speakers — the cleaner test — WER fell from 22.44% (best prompt) to 12.17%, a 45.8% relative reduction, beating an open medical-specialist model (13.79%) and a widely used general-purpose ASR (14.92%) on the same scorer. The whole adaptation took about 14 GPU-hours on a single GB300.

The trade-off is real but small: general English LibriSpeech test-clean WER rose from 1.54% to 1.69% after domain tuning. Prompts with term lists deliberately include near-miss distractors, so the model learns to trust a spelling only when the audio supports it.

If you run transcription workflows on meeting recordings, calls, or lectures, this is worth a local trial: the model card ships recommended presets and the full prompt guide, one-GPU serving is enough, and the open data pipeline means domain adaptation no longer requires vendor access. Spot-check names and speaker changes against the recording, as NVIDIA itself advises.

Watch for third-party leaderboard entries to firm up the vendor-reported accuracy figures. For workflow ideas around open models like this, see LinkLoot's guide to AI agent tools.

Source check

Primary evidence is NVIDIA's official announcement on the Hugging Face blog, with the published model card and dataset as artifacts. Baseten, an independent inference provider, separately lists and serves the model under nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, confirming production availability. Artificial Analysis is the independent ranking body for the AA-WER claims cited by NVIDIA; the specific scores remain vendor-reported pending published entries.