NVIDIA Releases Nemotron 3.5 Transcribe and Granary v2 for Recorded-Audio Transcription and Speaker Labels

NVIDIA published Nemotron 3.5 Transcribe on October 28, 2026, in a Hugging Face blog post that also introduces the Granary v2 speech dataset and fine-tuning recipes. Transcribe is an open-weight model that works on recorded audio, not live streaming, and it is a different product from the streaming speech model covered here on October 27. One checkpoint does transcription, speaker labels, translation and audio question answering, steered by a text prompt. The leaderboard results in the post are NVIDIA's reading of third-party rankings.

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. A speech encoder feeds a hybrid Mamba–Transformer mixture-of-experts language model, so only about 4B of its 30B parameters are active. Prompts can ask for a transcript with speaker labels, a translation into English or an answer such as which decisions a meeting made. Audio question answering is trained in English, and language support differs by task, so the model card's task-by-language matrix matters. NVIDIA advises spot-checking names, speaker changes and answers against the recording.

The leaderboard claims use different error measures, and they should not be merged. Word error rate (WER) is substitutions plus deletions plus insertions, divided by the number of words in the reference transcript. On the Hugging Face Open ASR Leaderboard, NVIDIA reports first place among open-weight models for English at a 5.33% average WER, with API models excluded. In the model card's comparison it lists a 4.62% mean WER across seven languages and 10.98% for long-form audio, both the lowest among the open-weight models evaluated there. Artificial Analysis uses its own AA-WER index: Transcribe ties for first among open-weight models at 2.8%, with about 4B active parameters against 24B for the model it ties.

Speaker results use other measures. On David AI's 21-language benchmark, diarization error rate (DER), which counts speaker-time errors, is 18.26%, the lowest of 14 systems and slightly ahead of NVIDIA's own Nemotron 3 Diarization at 18.38%. On VoiceArena's English full-session test, the concatenated minimum-permutation WER (cpWER), which counts word errors after matching speaker labels, is 32.4%, the lowest of 14 systems. Separately, deletions are 8.32% of reference words, against a median of about 31%. A deletion rate is not a WER. NVIDIA also says its forthcoming NVFP4 quantized version processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, measured on 30-second clips at high concurrency. That figure is not for the checkpoint released today.

Granary v2 is the open data pipeline behind Transcribe's training. Its public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in the first version; NVIDIA stresses these figures describe the dataset release, not Transcribe's full training mixture. The pipeline has four stages. Audio preparation segments audio, tracks speakers and tags language and acoustic events, and this release skipped it because its sources were already segmented. Transcription sends each language to primary and recovery recognizers and falls back when a hallucination check fails. Task-label generation derives transcript variants and checks character rate, hallucination patterns and agreement. A high-fidelity stage then removes noisy, clipped or multi-speaker segments; for VoxPopuli it cut 354k hours to 111k across 22 European languages.

The post's medical case study uses Eka Care's open benchmark of 3,619 clips from Indian doctors and medical students. Naming English and the Latin alphabet in the prompt cut full-benchmark WER from 30.8% to 23.8%, and fine-tuning brought it to 11.8%, about 62% fewer errors than out of the box. The cleaner test is the 2,764 held-out clips, where WER fell from 22.44% for the best prompt to 12.17%, a 45.8% relative reduction. The adaptation used about 153 hours of synthetic speech, 29 hours of accented English and 63 hours of medical recordings. The full process, from text generation through evaluation, took about 14 GPU-hours on one GB300. These are NVIDIA's results on one benchmark. General English got slightly worse, with LibriSpeech test-clean WER rising from 1.54% to 1.69%, and a medical specialist model still led on exact brand-name matches.

Teams that need recorded-call, meeting or dictation transcripts now have an open model with prompt control, plus data tooling to adapt it. The numbers are strongest as a reason to test, not a promise for any one workload. NVIDIA itself advises keeping at least 100 real, labeled clips for checkpoint selection and checking the entity types an application depends on.