NVIDIA releases Nemotron 3.5 Transcribe for instruction-led speech recognition

NVIDIA announced Nemotron 3.5 Transcribe on October 28, releasing an instruction-controlled speech model for turning recordings into transcripts, translations and speaker-labeled text. The open-weight release gives developers another option for processing audio themselves rather than relying entirely on a proprietary transcription service.

The release is available, not simply a research preview. NVIDIA’s launch article links downloadable weights and deployment tools. Its model card dates the launch to October 28, although the repository history includes October 27 preparation commits. A repository timestamp alone does not establish when unrestricted public access began.

More control over the transcript

The central idea is that transcription is not one fixed task. A meeting archive may need speaker labels, while a searchable record may favor normalized numbers and cleaner wording. NVIDIA’s technical report describes a speech encoder and speaker-detection component feeding a language model, allowing instructions to change the output convention without retraining.

NVIDIA says the system can handle recordings up to an hour long and transcribe 40 languages. That does not mean every feature supports all 40: the launch article specifies translation between English and 31 languages, speaker attribution in 13, and English training for audio question answering. The model is for recorded audio, not a live-streaming voice assistant.

For developers, the useful distinction is control over both processing location and transcript format. Those choices could simplify meeting archives or specialized transcription workflows, but they do not eliminate the need to check names, overlapping speech and answers against the original recording.

Available weights, but deployment claims need care

The model card identifies a 34-billion-parameter system with roughly 4.3 billion active parameters, including its audio encoder, under the OpenMDW-1.1 license. Its reference deployment uses an H100 with 80 GB of memory. The active-parameter figure should not be read as the total memory footprint or a promise that an ordinary laptop can run the reference checkpoint.

NVIDIA’s release materials are not fully synchronized: the launch article describes quantized versions and some local deployment options as forthcoming, while the model card and technical report list quantized variants. The verified downloadable reference checkpoint is the safer starting point; developers should check the exact variant and supported hardware before planning a deployment.

The card also separates its fastest serving results from the released development stack. Its headline throughput uses optimized NIM serving on specific Blackwell hardware; NVIDIA says those optimizations are not yet in the open-source release. Those numbers should not be treated as the performance of every downloadable configuration.

What the benchmarks establish—and what they do not

The technical report reports a 4.71% mean word error rate across nine English test sets and 4.62% across seven multilingual sets. These are NVIDIA-reported evaluations, not tests performed by PromptsZone. Word error rate counts substitutions, omissions and insertions; it is not a universal accuracy percentage for meetings, accents or specialist terminology.

The practical test is whether a chosen checkpoint improves a representative set of recordings without losing important words or assigning them to the wrong speaker. Keep separate evaluation recordings, compare both transcript quality and operating cost, and retain human review where errors have consequences. Open weights make that evaluation more controllable—not unnecessary.

Photo: Jensen Huang at Stanford, April 30, 2026, by Anders Eidesvik / Wikimedia Commons, CC BY-SA 4.0. Unmodified archival company-leadership photograph, not the Transcribe launch.