Key Notes
- Nemotron 3.5 Transcribe processes recordings up to an hour long in 40 languages, including Russian, while audio question answering has been evaluated only in English.
- NVIDIA reports leading open-weight results on selected speech benchmarks, rather than a guarantee of better transcription on every recording.
- The reference deployment uses an H100 with 80GB of memory, and developers should verify the availability and hardware support of lower-precision variants.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that turns recordings of up to an hour into text across 40 languages, including Russian. It can also add speaker labels and word timestamps, translate speech and answer questions about recorded audio, although language support varies by task.
The October 28 release is available on Hugging Face, where NVIDIA’s model card sets out its capabilities and deployment requirements. The focus is processing finished recordings: interviews, meetings and calls that developers want to turn into useful information after they happen.
One Recording, Several Useful Outputs
A conventional transcript does not suit every job. A researcher may want every hesitation preserved, while someone preparing meeting notes may prefer cleaner wording and numbers written as digits. NVIDIA’s research overview describes a model that takes those instructions in ordinary language, allowing developers to change the output style without retraining it.
The same approach extends to specialist vocabulary: a request can include expected names or product terms to help the model interpret the recording. Speaker attribution separates contributions within a conversation, while word timestamps make individual passages easier to find. Those speaker labels distinguish voices in the recording; they do not establish the real-world identity of each person.
According to NVIDIA’s launch breakdown, translation covers 31 languages to or from English, and speaker-attributed transcription covers 13 languages with up to eight speakers. Russian is listed for transcription, English translation and multi-speaker processing. That does not mean every feature has been validated equally across all 40 languages.
Audio question answering is a particularly important distinction. NVIDIA says its audio Q&A, reasoning and spoken-instruction capabilities have been evaluated only in English. Russian transcription support should therefore not be read as a promise that questions about a Russian recording will work just as reliably. The model is also built for recorded audio, rather than a live streaming conversation.
What NVIDIA’s Accuracy Results Show
In its technical report, NVIDIA reports a mean word error rate of 4.71% across nine public English test sets, compared with 4.94% for Qwen3-ASR-1.7B, the next open-weight model in that comparison. Lower word error rates are better, but the figure does not mean that 95.29% of complete transcripts are error-free.
The reported averages are 10.98% on the long-form track and 4.62% on a multilingual benchmark covering seven languages. NVIDIA describes those as leading open-weight results within the comparisons it publishes. Its own report also shows commercial systems ahead on some aggregates and other open models ahead on individual datasets, so this is not evidence of universal superiority.
For a team choosing transcription software, the practical test remains its own recordings: accents, overlapping voices, background noise and specialist terminology can change the outcome. A clean benchmark average is useful evidence, but names, amounts and consequential statements still deserve a check against the original audio.
Running It Locally Still Takes Serious Hardware
NVIDIA’s reference setup uses an H100 GPU with 80GB of memory for the BF16 checkpoint. That is a documented starting configuration, rather than a requirement to buy that exact card: the compatibility list includes other hardware, with support depending on the model’s numerical precision and software stack.
There is a wrinkle for developers looking for smaller downloads. The model card lists FP8 and NVFP4 variants, while NVIDIA’s October 28 release update describes quantized versions as coming soon. Check the actual checkpoint and supported configuration before assuming that a compact version is ready for a particular desktop.
The weights are released under OpenMDW 1.1, with NVIDIA describing commercial and noncommercial use as supported. Running them still carries hardware or hosting costs. Alongside its Nemotron agent models, NVIDIA is expanding the building blocks developers can operate themselves; transcription now joins a market that also includes a competing speech model from Google. The appeal is more control over where recordings are processed and how the resulting text is prepared.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
