Speech stacks are usually a pipeline: an ASR model writes words, a separate diarization model labels who said them, a translation model rewrites the result. Each stage loses information and each stage needs its own hosting, tuning and error handling. NVIDIA’s Nemotron 3.5 Transcribe collapses that into a single open-weight checkpoint that does all of it in one generation pass.
The model (30B-A4B, a hybrid Mamba–Transformer mixture-of-experts design where roughly 4B parameters are active per token) transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. Recordings up to an hour long are supported, with up to eight distinct speakers labeled. It also handles audio question answering — ask “what decisions were made in this meeting?” and get an answer grounded in the audio, not just a transcript to search.
The numbers
On the Hugging Face Open ASR Leaderboard, Transcribe is first among open-weight models on English at 5.33% average WER; multilingual sits at 4.62% mean WER across seven languages and long-form at 10.98%. On Artificial Analysis it ties for first among open-weight models at 2.80% AA-WER while carrying about 4B active parameters against 24B for the model it ties.
Speaker attribution is where the model pulls ahead of dedicated diarization systems. On David AI’s 21-language leaderboard it posts a 18.26% diarization error rate, narrowly ahead of NVIDIA’s own dedicated Nemotron 3 Diarization model at 18.38%. On VoiceArena’s 14-system comparison it leads with 32.4% cpWER and deletes far fewer words than the median — 8.32% of reference words against roughly 31%. The trick is that Sortformer speaker cues are fed into the language model alongside acoustic features, so speaker turns are decided during generation instead of being stitched on afterwards.
Throughput was benchmarked at 100 hours of audio in under three minutes on a single RTX PRO 6000 Blackwell Server Edition, using the forthcoming NVFP4 quantized build. Quantized weights and NeMo-Speech.cpp local deployment are not out yet, but vLLM serving works today.
The open dataset behind it
Granary v2, published alongside the model, is the more durable artifact: 463,062 hours across 48 languages and 16 source corpora, built with NVIDIA NeMo Curator into aligned transcript variants, task labels and inspectable quality metadata. For a promptable speech model to follow formatting instructions, training targets have to be unambiguous — and Granary v2 supplies that at a scale most speech labs keep private.
The fine-tuning story is proven rather than promised. In a medical case study, adapting the model cut WER on held-out speakers by 45.8% relative to the best prompt — the difference between telling the model “HbA1c, metformin” and retraining it.
Related: Nemotron 3.5 Transcribe model card, Granary v2 dataset