NVIDIA has announced Nemotron 3.5 Transcribe, an open-weights speech-augmented language model designed for audio transcription. According to NVIDIA’s technical report and model information, the model supports 40 languages and audio recordings of up to one hour per request. Text instructions can request formats including verbatim transcription, written-form output, speaker-attributed transcripts, contextual transcription and translation.

The checkpoint contains about 34 billion total parameters, with approximately 4.3 billion active parameters, and has a 65,536-token context. NVIDIA lists BF16, FP8 and NVFP4 versions. The company says the model can run in BF16 on one H100 80 GB GPU, while the NVFP4 version runs on an RTX 5090. Supported runtimes include vLLM with the NeMo Speech plugin, Hugging Face Transformers, SGLang-Omni and NeMo-Speech.cpp.

NVIDIA reports a 4.71% mean word error rate across nine public English Open ASR datasets and 4.62% across seven multilingual datasets. The figures are self-reported; NVIDIA also states that commercial systems achieved lower error rates on the cited benchmarks. The report identifies weaker results on Mandarin far-field meeting datasets, including a 34.95% character error rate on AliMeeting.

The model combines a trainable ASR encoder with a diarization component supporting up to eight speakers. A separate alignment adapter provides word timestamps. NVIDIA says an entity list reduced named-entity word error rate on ContextASR-Bench from 12.21% to 2.20%.

NVIDIA describes the release as offline rather than a streaming ASR service. Its OpenMDW-1.1 license covers the model release, while some serving optimizations are not open source.