AI Industry

NVIDIA Releases Open Nemotron 3.5 Transcribe Speech Model

Oct 28, 2026·4 min read·By Rahish Saifi

Unbranded black conference puck microphone on a sunlit oak meeting table beside a notebook and pen, the headline NVIDIA OPENS SPEECH AI with the orange subline Transcribes 40 languages, and an AIWithRahish badge in the bottom-right corner

On Oct. 28, 2026, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech LLM for promptable transcription, alongside Granary v2 data and curation tools. NVIDIA's research page and Maryam Motamedi's Hugging Face post frame the drop for business meetings, medical dictation, and multilingual calls—settings where transcript format and speaker labels matter as much as raw words.

Key facts

Item	Detail
Release	Oct. 28, 2026 — NVIDIA research page, Hugging Face / ModelScope weights, Granary v2 manifests
Architecture	Speech encoder + frozen Nemotron 3 Diarization + Nemotron 3.5 Lightning 30B-A3B MoE LLM; 34B total, ~4.3B active (model card)
Capabilities	Multiple transcription styles, translation, language ID, audio Q&A via natural-language instructions—no retraining per style
Languages	40 transcribed; 31 translated with English; diarization in 13; ≤8 speakers; ≤1 hour audio; 65,536-token context
Open release	OpenMDW-1.1; BF16/FP8/NVFP4/GGUF + DFlash2; 1× H100 80GB; NeMo Speech fine-tuning
Granary v2	48 languages; manifest 463,062 h / ~843,600 h processed; open NeMo Curator pipeline

What NVIDIA released

NVIDIA describes Nemotron 3.5 Transcribe as a speech-augmented language model (SALM): a Transformer speech encoder, a frozen Nemotron 3 Diarization model whose speaker activity is injected into encoder states, and the Nemotron 3.5 Lightning 30B-A3B hybrid Mamba–Transformer MoE LLM. The model card lists 34B total parameters and ~4.3B active, including the audio encoder.

Weights use the OpenMDW-1.1 license on Hugging Face and ModelScope (BF16, FP8, NVFP4, GGUF, DFlash2 draft). NVIDIA also ships NeMo Speech fine-tuning tools and published evaluation prompts.

What one model can do

One checkpoint covers modes that usually need separate ASR, diarization, and post-processing stacks: verbatim, written-form, contextual (entity lists cut named-entity errors on ContextASR-Bench six-fold, per NVIDIA), speaker-attributed, and rich transcription, plus speech translation, spoken language ID, and audio Q&A. Motamedi's examples—label speakers, translate to English, or summarize decisions—use the same weights with different instructions; the model card maps which tasks run in which languages (audio Q&A is English-trained).

Speaker cues from the frozen diarization model feed the encoder so overlapping meetings can be transcribed with tags during generation. NVIDIA claims state-of-the-art speaker-attributed ASR from one-minute clips through full calls and says it is, to its knowledge, the first SALM to beat specialized cascades on that task. A lightweight CTC adapter can then align text to audio for speaker-attributed word-level timestamps on request—useful for replay UIs such as Agora's classroom demo.

Languages and limits

Transcription covers 40 languages. Translation supports 31 languages to or from English. Speaker attribution works in 13 languages. The model accepts up to eight speakers, handles recordings up to one hour, and exposes a 65,536-token context window. Word-level timestamps are available when requested in the prompt.

NVIDIA and Motamedi stress recorded, offline audio—not live streams. TechCrunch's Oct. 28 coverage notes the same limitation for builders comparing Transcribe to realtime agent stacks such as Mistral Voxtral 3 or speech-to-speech systems like xAI Grok Voice 3.

Specs and how to run it

Checkpoints ship in BF16, FP8, NVFP4 (Blackwell), GGUF for NeMo-Speech.cpp, plus an optional DFlash2 draft for speculative decoding alongside joint multi-token prediction in vLLM. The model card lists 1× H100 80GB deployment; NVIDIA's research page cites up to 2,238 seconds of audio transcribed per second with NVFP4 and DFlash2 on one RTX PRO 6000 Blackwell GPU via the Nemotron Transcribe NIM.

Motamedi points builders to vLLM or SGLang-Omni for self-hosting, or hosted endpoints from Baseten, DeepInfra, and Simplismart. NVIDIA claims 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition with the forthcoming NVFP4 build (30-second clips, high concurrency). Quantized local packages for Spark, Thor, and NeMo-Speech.cpp were flagged as coming soon at launch.

Granary v2

Granary v2 is the open Nemotron speech collection behind Transcribe, curated with NeMo Curator. Its public manifest covers 463,062 hours in 48 languages and 16 corpora (up from 25 languages in v1); NVIDIA puts the full processed collection near 843,600 hours. Each clip can include verbatim, disfluent, and cleaned written forms—the same targets you can ask for at inference—plus captions, code-switch fixes, speech Q&A pairs, biasing terms, and scored translations.

NVIDIA open-sources the pipeline so teams can rebuild manifests on authorized audio, then fine-tune in NeMo Speech. Motamedi's medical case study reports a 45.8% relative held-out WER drop after fine-tuning versus the best prompt-only baseline on Eka Care's Indian medical speech benchmark.

Why it matters

Analysis. Most open speech models still commit to one transcript style or chain ASR, diarization, and cleanup. NVIDIA is betting a promptable SALM with diarization injected into the encoder can serve archived meetings, dictation, and compliance audio from a single weight file—then adapt through instructions or NeMo fine-tuning when prompts are not enough.

NVIDIA claims first-place open-weight results on Hugging Face's Open ASR Leaderboard (Motamedi cites 5.33% average English WER; NVIDIA's research page lists 4.71% on nine English sets and 4.62% multilingual), plus leading open-weight scores on Artificial Analysis, David AI, and VoiceArena diarization tests—claims about offline accuracy, not live agent latency. Pairing those weights with Granary v2 manifests makes the release harder to dismiss as a one-off checkpoint: the formatting targets and curation code are part of the story.

What to watch

Teams should verify leaderboard numbers on their own audio—Motamedi advises spot-checking names, speaker swaps, and Q&A answers. Delivery of NVFP4, GGUF, and NeMo-Speech.cpp builds will show how quickly “run on Spark/Thor” becomes real. Transcribe will not replace realtime stacks like Voxtral 3, but richer offline transcripts may feed those agents; watch domain fine-tuning recipes (NVIDIA cites ~0.5–2.3 GPU-hours per task on 17 adaptations) for how fast medical and legal verticals fork the base model.