NVIDIA releases open-weight Nemotron 3.5 Transcribe for 40 languages

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model. It accepts up to one hour of audio and takes a plain-language instruction describing the transcript style. It handles 40 languages, with optional speaker labels and word-level timestamps. The NVIDIA technical report is dated October 28, 2026, and the model is released under the OpenMDW-1.1 license.

The report starts from the premise that a transcript has no single correct form. Whether numbers should be spelled out or written as digits, whether filler words stay or get cleaned up, how domain terms are notated, and which script is used for code-switched speech all depend on who will read the transcript and why. Conventional recognizers, NVIDIA argues, commit to one of these conventions during training. Nemotron 3.5 Transcribe chooses the convention per request, through an instruction, without retraining.

The system is a speech-augmented language model, or SALM. A Transformer speech encoder and a frozen Nemotron 3 Diarization model feed into the Nemotron 3.5 Lightning 30B-A3B LLM, a hybrid Mamba-Transformer mixture-of-experts model. Frozen means the diarization model's weights are not updated during training. Its speaker-activity signals are injected into the encoder's representations, which lets the language model produce speaker-attributed text without changing its own architecture. The report says the LLM has about 4.3B active parameters, but the checkpoint names read 30B-A4B, and the report does not reconcile the two figures. Word timestamps come from a separate lightweight CTC adapter. It aligns the finished transcript to the audio after decoding, which NVIDIA says works for long and heavily overlapping recordings.

NVIDIA's headline accuracy claim is a mean word error rate of 4.71% across the nine public test sets of the English Open ASR Leaderboard, and 4.62% on the multilingual version. Word error rate counts substitutions, deletions and insertions, so lower is better. NVIDIA calls this the best open-weights result on those leaderboards. The speaker-attributed claim is more specific. The report says the model reaches state-of-the-art accuracy on overlapping multi-speaker speech, from one-minute segments to full meetings and calls. It also says the model is the first SALM to outperform specialized cascaded systems on that task, to the authors' knowledge. A further claim is that, when given a list of expected entities, the model cuts named-entity errors six-fold on ContextASR-Bench. The excerpt gives no baseline for that six-fold figure. All of these are results NVIDIA reports, and the excerpt does not break accuracy down by language, so the 40-language coverage is not shown language by language.

The throughput figure depends on a particular setup. The report says the NVFP4 checkpoint, paired with the DFlash2 draft model, transcribes up to 2,238 seconds of audio per second on a single RTX PRO 6000 Blackwell GPU, served through the Nemotron Transcribe NIM. Speculative decoding works by having a small draft model propose several tokens that the large model then checks, which can speed generation when most proposals are accepted. The 2,238 figure applies to the NVFP4 and DFlash2 configuration, not to the full-precision BF16 checkpoint, and the excerpt does not give throughput for the other formats.

NVIDIA publishes BF16, FP8, NVFP4 and GGUF checkpoints, plus the DFlash2 draft model. NVFP4 is aimed at Blackwell GPUs, and the GGUF files target NeMo-Speech.cpp. A fine-tuning toolkit ships in NeMo Speech. The report says the model can be adapted to new languages, domains and tasks on a single GPU, typically within 0.5 to 2.3 GPU-hours per task, and that this was shown on 17 tasks.