NVIDIA releases open-weight Nemotron 3.5 Transcribe speech model

NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech recognition model built to handle the fact that there's no single correct way to transcribe audio. The model pairs a speech encoder and a frozen speaker-identification model with NVIDIA's Nemotron 3.5 Lightning language model, for about 4.3 billion active parameters, and can produce verbatim text, cleaned-up written form, speaker-labeled transcripts, translations, and language identification across 40 languages from a single model, chosen with a plain-language instruction rather than retraining.

NVIDIA says it is the best open-weights model on the Open ASR Leaderboards, with a mean word error rate of 4.71% on the English leaderboard's nine public test sets and 4.62% on the multilingual leaderboard. The company also claims it's the first speech-augmented language model to beat specialized systems built specifically for multi-speaker transcription, and that it cuts named-entity errors six-fold when given a list of expected terms. On a single RTX PRO 6000 Blackwell GPU, NVIDIA says an optimized version can transcribe up to 2,238 seconds of audio per second.

Artificial Analysis separately reported, in its own benchmark, that the model ties for first place among open-weights speech-to-text models with a 2.8% word error rate. NVIDIA released the model, along with a fine-tuning toolkit, under the OpenMDW-1.1 license.

Why it matters

An open-weight model that handles multiple transcription styles and 40 languages in one system lowers the cost of building voice products without relying on closed APIs.

Source: NVIDIA Research