October's speech-transcription lane is crowded: Mistral shipped Voxtral 3 in late October, IBM updated Granite Speech 5 in the same week, and Gemini 3.5 Transcribe and Muse Voice Transcribe landed in August and September. On October 28 NVIDIA entered with three things at once: the open-weights speech LLM Nemotron 3.5 Transcribe, the 463,062-hour training-data pipeline Granary v2, and a fine-tuning toolkit.

There is no single correct transcript

The technical report opens with its premise: a recording has no single correct transcript — verbatim or written form, "twenty-five euros" or "€25", fillers kept or dropped — it depends on who reads it and for what purpose. Conventional ASR locks that convention in at training time; here the convention becomes a prompt. Transcription style, domain vocabulary, and speaker-tag format are chosen per request in natural language, with no retraining. Granary v2 renders every training recording into at least three transcript variants (verbatim with fillers, written form with fillers, clean written form), so the model learns to follow a requested style instead of guessing one.

Architecturally it is a SALM: a Transformer speech encoder plus a frozen Nemotron 3 Diarization branch (a Sortformer, about 100M parameters, 8 speaker slots) injects who-spoke-when activity into the acoustic representations, feeding a Nemotron 3.5 Lightning 30B-A3B hybrid Mamba-Transformer MoE backbone — roughly 34B total and 4.3B active parameters. One model covers transcription in 40 languages, translation between English and 31 of them, attribution for up to 8 speakers with annotation in 13 languages, and recordings up to an hour long.

Strong numbers, except Mandarin meetings

Officially: a mean WER of 4.71% across the nine public test sets of the English Open ASR Leaderboard and 4.62% on the multilingual board, claimed best among open-weights models; Chinese tech outlet IT Home reports it ranks first among open-weights models on the English board. On the third-party side, per IT Home: 2.8% on Artificial Analysis' non-streaming WER index, tied with Voxtral Small Transcribe at the top of open models, and the lowest 18.26% error on David AI's 21-language speaker-attribution benchmark. For overlapping multi-speaker speech the team claims state-of-the-art accuracy and, to its knowledge, the first SALM to beat specialized cascaded systems; given a list of expected entities, ContextASR-Bench named-entity errors drop six-fold.

The same results table carries the cold water: full English meetings land at 12.91% cpWER, while Mandarin meetings on AliMeeting still sit at 34.95% cpCER. Multi-speaker Mandarin remains the hardest bone to chew.

2,238x real time, and it fits locally

In NVIDIA's own benchmark, the NVFP4 checkpoint with the DFlash2 draft model transcribes up to 2,238 seconds of audio per second on a single RTX PRO 6000 Blackwell. The model's own multi-token-prediction layer doubles as the speculative-decoding draft in vLLM. All four precision variants — BF16, FP8, NVFP4 and GGUF — are open, with GGUF paired to NeMo-Speech.cpp; IT Home reports builds for local devices such as the DGX Spark and Jetson AGX Thor will follow. Fine-tuning is cheap: 17 tasks each took 0.5-2.3 GPU-hours on a single GB300, and 13 minutes of Swahili audio cut WER from 73.7% to 40.8%. In a medical case study, about 150 hours of synthetic speech plus real accented data brought an Indian-English medical benchmark from 24.03% down to 12.06% — better than Whisper-large-v3 at 14.92% — in roughly 14 GPU-hours.

Transcription quality is agent cost

The appendix's most persuasive experiment is a voice-agent run: across 100 airline customer-service scenarios, the agent completed 58% of tasks with this recognizer, at 8.6% WER on user speech versus 12.9% for Parakeet-CTC-1.1B. Tool-call failures caused by misrecognized surnames fell from 73 to 2, and each interaction saved 2.7 turns, about 37,000 LLM tokens, and 78 seconds. Every point of WER you shave is a detour the agent no longer takes — a more concrete account than any leaderboard rank.

The limits are stated plainly: preserving the base LLM's text-only abilities is "explicitly not a goal" — this is a transcription specialist, not a general assistant; PII redaction is still a small, synthetic, English-only probe; and 34B total parameters is not exactly light for the edge. But the bundle — a model plus a 463k-hour data pipeline plus tooling — is a platform play: transcription is shifting from single-model races to stack wars. The next entrant without a data pipeline may not even bother holding a launch event.