NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, an open speech model that transcribes recordings in 40 languages, tells speakers apart and translates audio. The news is less about plain speech recognition than about the approach: instead of a fixed output format, developers use a text instruction to decide what the transcript should look like, whether verbatim with filler words, cleaned up, with speaker labels or translated straight into English. The weights are available on Hugging Face, and the license allows commercial use.
That matters for teams processing meetings, support calls or dictation, because many of them still chain several systems together: one for recognition, one for speaker attribution, one for translation. Here is what the model can do, which numbers NVIDIA reports and where the limits are.
Key Takeaways
- Nemotron 3.5 Transcribe transcribes 40 languages, translates 31 of them to and from English and attributes speakers in 13, for recordings of up to one hour and up to eight speakers.
- It is a speech-augmented language model with roughly 4.3 billion active parameters; the model card lists 34 billion parameters in total, including the audio encoder.
- NVIDIA lists German, French, Hindi, Korean and other languages at its highest quality tier (“Leading”).
- License: OpenMDW-1.1, which permits commercial and non-commercial use.
- Limitation: the model is built for recorded audio, not live streaming. Audio question answering is trained in English only.
What is Nemotron 3.5 Transcribe?
According to the model card on Hugging Face, Nemotron 3.5 Transcribe is an offline model that takes audio recordings together with text instructions and optional context and returns text. NVIDIA calls it a speech-augmented language model (SALM). Under the hood it couples a Transformer speech encoder and the frozen Nemotron 3 Diarization model with the Nemotron 3.5 Lightning 30B-A3B language model, a hybrid Mamba-Transformer with a mixture-of-experts design. NVIDIA says about 4.3 billion parameters are active per token.
The context window is 65,536 tokens, and recordings can be up to an hour long. Besides transcription, NVIDIA says the model handles speech translation, language identification, contextual transcription with supplied domain terms, word timestamps and questions about the audio. Weights ship in BF16 as well as FP8 and NVFP4 variants. NVIDIA’s Nemotron 3 family recently drew attention with results at academic olympiads in math and informatics; now a specialized speech model joins the lineup.
Steering transcription with instructions
The core idea is control. In its announcement post, NVIDIA argues that there is no single correct transcript: meeting minutes need different conventions than medical dictation or subtitles. Classic speech recognition commits to one convention during training. Nemotron 3.5 Transcribe instead receives it with each request as a natural-language instruction, such as “Transcribe this recording and label each speaker” or “Translate this audio into English.”
For domain vocabulary, NVIDIA recommends passing terms directly in the prompt, with a note to use them only when the audio supports them. On ContextASR-Bench, such a term list cuts named-entity errors six-fold, NVIDIA writes on the research page for the model. On request, the model also returns word timestamps with speaker labels, which helps editing and review workflows. An experimental prompt is meant to transcribe only the person nearest to the microphone and ignore background voices.
Nemotron 3.5 Transcribe in the benchmarks
NVIDIA points to several public leaderboards. Keep in mind that all figures come from NVIDIA’s own presentation; independent real-world tests are still pending.
- Open ASR Leaderboard (English): first among open-weight models. The blog post cites a 5.33% average word error rate on the leaderboard, while the model card reports 4.71% as the equal-weight mean across nine public test sets (captured October 12). The figure depends on how it is calculated. The next-best open model, per the model card, is Qwen3-ASR-1.7B at 4.94%.
- Multilingual leaderboard: 4.62% mean word error rate across seven languages, versus 6.49% for the next-best open model.
- Long-form audio: 10.98% mean word error rate, compared with 13.37% for Parakeet-TDT-0.6B-v3.
- Artificial Analysis: tied for first among open-weight models at 2.8% AA-WER, with about 4 billion active parameters versus 24 billion for the model it ties.
- Speaker attribution: the lowest diarization error rate (18.26%) on David AI’s 21-language leaderboard, and first place in VoiceArena’s 14-system comparison at 32.4% cpWER.
Language support varies by task. For German, for example, the model card lists transcription, translation, multi-speaker recognition and code-switching at the “Leading” tier, with a 2.73% word error rate on FLEURS and 2.98% on CoVoST2.
Hardware, license and deployment
As a deployment reference, the model card names a single 80 GB H100. Supported GPUs span the Ampere, Hopper, Lovelace and Blackwell generations depending on precision, including A100, H100, L40S, DGX Spark and RTX 5090, the latter with NVFP4 only. At launch, the model runs with vLLM 0.32.0 and the NeMo Speech plugin, Hugging Face Transformers and SGLang-Omni. Teams that do not want to self-host can use hosted endpoints from Baseten, DeepInfra and Simplismart. NVIDIA says local deployment via NeMo-Speech.cpp is coming soon.
Speed claims deserve a look at the fine print. NVIDIA cites up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell Server Edition. That figure applies to its own inference service (NIM) with NVFP4 and speculative decoding. According to the model card, these optimizations are not yet part of the open-source release, and the bundled vLLM integration is expected to be roughly half as fast.
The OpenMDW-1.1 license permits commercial and non-commercial use. As with any license, review the terms yourself before production use, especially since third-party components may carry their own notices.
Granary v2 and fine-tuning
Alongside the model, NVIDIA published Granary v2, the data pipeline behind part of the training. NVIDIA says the public manifest covers 463,062 hours across 48 languages from 16 source corpora, up from 25 languages in Granary v1. Those figures describe the dataset release, not the full training mix. Each recording comes with three transcript forms, from verbatim with fillers to cleaned up with numerals.
NVIDIA illustrates the value of adaptation with a case study on medical recordings from Indian doctors and medical students. On speakers held out from tuning, fine-tuning cut the word error rate from 22.44% with the best prompt to 12.17%, a 45.8% relative reduction. The whole adaptation took about 14 GPU-hours on one GB300, according to NVIDIA. On general English, the error rate rose slightly, from 1.54% to 1.69%.
Limits and open questions
NVIDIA is fairly explicit about the limits. Nemotron 3.5 Transcribe is an offline model; processing audio internally in 30-second chunks does not make it a real-time service, so it is not meant for voice assistants with live audio. Audio understanding, reasoning and spoken instructions were evaluated in English only and may not work reliably in other languages. And as with any speech model, NVIDIA advises spot-checking names, speaker changes and answers against the recording.
Companies in Europe face another question: anyone processing recorded conversations is almost always processing personal data. An open model running on in-house hardware can be an advantage over cloud APIs here, but it does not replace a privacy review. At the same time, the model keeps NVIDIA closely tied to its own hardware, much like its recent AI partnership with Germany: open weights are also a way to sell GPUs.
Bottom line
Nemotron 3.5 Transcribe bundles transcription, speaker attribution, translation and audio questions into one open model that can be steered with instructions. The benchmark numbers are strong, but they come from NVIDIA’s own selection and vary depending on how they are calculated. For developers who want to process recordings on their own infrastructure, in English, German or many other languages, the model is worth testing. Anyone who needs live transcription or only has modest hardware will have to stick with other options or wait for the announced local variants.
