NVIDIA open-sources Nemotron 3.5 Transcribe with a promptable speech stack

NVIDIA used its Hugging Face blog on October 28 to publish Nemotron 3.5 Transcribe, an open-weight speech model that ships with matching data and fine-tuning tooling. The checkpoint is available on Hugging Face and ModelScope, and it consolidates four jobs that usually require separate systems: transcription, speaker attribution, translation and audio question answering.

The model transcribes 40 languages, translates 31 of them to or from English, and labels speakers in 13 languages. It accepts recordings up to an hour long and can tag as many as eight speakers. Output is steered with plain-language instructions: a developer can ask for a transcript with speaker labels, an English translation, or a summary of the decisions made in a meeting, all from the same weights.

Architecturally, Transcribe pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model. Roughly 4 billion of its 30 billion parameters are active per token, which keeps inference cheap enough for a single GPU, from Jetson AGX Thor and DGX Spark up to Blackwell data-center parts. NVIDIA says one RTX PRO 6000 Blackwell Server Edition GPU at NVFP4 precision processes 100 hours of audio in under three minutes. It targets recorded audio rather than live streaming, and can be served through vLLM or NeMo-Speech.cpp.

The benchmark claims are specific. On the Hugging Face Open ASR Leaderboard, Transcribe posts 5.33% average word error rate in English, first among open-weight models, plus 4.62% mean WER across seven languages and 10.98% on long-form audio. Artificial Analysis records 2.8% AA-WER, tied for first among open-weight systems at roughly one-sixth the active parameters of the model it ties, and 1.8% WER on the held-out AA-AgentTalk voice-agent set against 2.3% for the next open model.

Speaker attribution is where the release departs most from convention. Sortformer speaker cues reach the language model alongside acoustic features, so speakers are tagged during generation rather than attached afterwards. On David AI's 21-language diarization track, Transcribe records 18.26% diarization error rate, narrowly ahead of NVIDIA's dedicated Nemotron 3 Diarization at 18.38%. On VoiceArena's 14-system English benchmark it leads at 32.4% cpWER, and drops 8.32% of reference words against a median near 31%.

Granary v2 is the data pipeline behind the model. Its public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1. For each recording the pipeline produces three transcript forms — text-normalized verbatim, inverse-normalized with disfluencies, and clean — the same styles a developer can request by prompt. It also generates audio captions, code-switching restoration, speech-grounded question-answer pairs and scored translations, built on NeMo Curator with Ray and Dynamo backends.

The accompanying case study shows why the tuning path matters. On Eka Care's 3,619-clip Indian medical speech benchmark, naming the language and script in the prompt cut full-test WER from 30.8% to 23.8%; fine-tuning brought it to 11.8%. On 2,764 held-out clips the adapted model reached 12.17% WER against 22.44% for the best prompt, a 45.8% relative reduction, beating an open medical specialist at 13.79% and a widely used general ASR model at 14.92%. The trade-off is a small regression on LibriSpeech test-clean, from 1.54% to 1.69%, and the whole adaptation ran in about 14 GPU-hours on one GB300.

The signal here is less about one leaderboard than about packaging. A single promptable checkpoint with open data manifests and a reproducible tuning recipe lowers the cost of building multilingual meeting, call-center and clinical transcription, and hosted endpoints from Baseten, DeepInfra and Simplismart cover teams that do not want to self-host. What to watch next is how quickly the community adapts it to new languages and code-switching, and whether NVIDIA extends the approach from recorded files toward the streaming workloads the model card currently excludes.