NVIDIA launches Nemotron 3.5 Transcribe with instruction-guided transcription

NVIDIA has introduced Nemotron 3.5 Transcribe, a speech-augmented language model that adapts transcription to instructions supplied with each request. It supports verbatim and written-form output, contextual transcription, speaker attribution, speech translation, and spoken language identification in 40 languages. The model handles recordings up to one hour long and can return speaker-attributed word-level timestamps, including for overlapping conversations.

NVIDIA reports a mean word error rate of 4.71% on the 9 public test sets of the English Open ASR Leaderboard and 4.62% on the multilingual leaderboard, describing it as the best open-weights model on those benchmarks. Supplying expected entities reduces named-entity errors on ContextASR-Bench six-fold. The system combines a speech encoder, a frozen diarization model, and a language model to support transcription that distinguishes speakers without changing the language model’s architecture.

The release includes model checkpoints and a fine-tuning toolkit in NeMo Speech under the OpenMDW-1.1 license. NVIDIA demonstrated adaptation across 17 tasks on a single GPU, typically within 0.5-2.3 GPU-hours per task. For inference, an NVFP4 checkpoint paired with the DFlash2 draft model and served through Nemotron Transcribe NIM transcribes up to 2,238 seconds of audio per second on a single RTX PRO 6000 Blackwell GPU.