Nvidia Nemotron 3.5 Transcribe Turns Recorded Speech Into Text in 40 Languages

Nvidia Nemotron 3.5 Transcribe in plain terms

Nvidia released Nemotron 3.5 Transcribe on October 28, a free open-weight model that turns recorded audio into text in up to 40 languages. It can also translate 31 languages to and from English, handle recordings up to an hour long, and tell up to eight speakers apart. If you have ever sat through a meeting recording trying to work out who said what, that last feature is the one that matters most.

The model is aimed squarely at developers building apps for business meetings and medical transcription, two places where voice AI already earns its keep. Nvidia says developers can pair it with the updated Granary speech dataset and tools like NeMo Speech to tune recognition for an industry's jargon in each language. A hospital that says "myocardial infarction" and a law firm that says "disclosure schedule" do not speak the same English, and a transcription model that cannot learn that difference will keep producing confident nonsense.

One honest limit: Transcribe works with recorded audio, not livestreams. If you need live captions for a call happening right now, this is not that tool. For recorded interviews, lectures, hearings, and meeting archives, the limit is less of a problem, and it keeps the design simpler.

Why open weights change the maths

Nvidia claims Transcribe beats rivals on benchmarks, including Hugging Face's Open ASR Leaderboard, and says the 30-billion-parameter model can run on a single chip. That range runs from the Jetson AGX Thor robotics controller and the DGX Spark chip in new laptops up to Blackwell datacenter chips. Running transcription locally matters for a simple reason: audio often contains the most sensitive things people say in a workday. Sending every meeting to a third-party API is a privacy and compliance decision, not just a technical one. Open weights let a company keep the audio inside its own walls.

That also changes cost. A per-minute API bill grows with every hour of audio you process. A model you host yourself costs hardware and engineering time up front, then scales more gently. For teams transcribing thousands of hours of support calls or clinical notes, the difference adds up fast. Nvidia's earlier Nemotron 3.5 retrieval models showed the same pattern: the unglamorous plumbing, search and now transcription, is where open models quietly save real money.

How Nemotron 3.5 Transcribe fits the voice race

Transcribe arrives in a busy month for voice AI. Mistral shipped Voxtral 3 earlier in October, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni. Each targets a different slice: live agents, real-time conversation, or recorded transcription. Nvidia is betting that recorded audio is still a huge market on its own, because most of the audio worth transcribing already exists as a file before anyone asks for a transcript.

The speaker separation feature deserves a closer look. Diarization, the technical name for working out who spoke when, is where many transcription tools fall apart in practice. A transcript that merges two speakers into one paragraph is hard to search, hard to quote, and useless as a meeting record. Eight-speaker support covers most real meetings, panels, and interviews. It will still struggle with crosstalk and bad microphones, because every model does, but the ceiling is higher than the two-speaker tools many teams still use.

What to try first

If you build with voice, start with your worst audio, not your best. A clean studio recording will flatter any model. Test a phone recording from a noisy room, a call with accents your team actually has, and a meeting where three people talk over each other. Check proper nouns, numbers, and medical or legal terms in your field, since those are where transcription errors cause real harm.

Also test the translation path separately from transcription. Translating 31 languages to and from English is a strong claim, and translation quality varies a lot by language pair. For high-stakes use, keep a human in the loop for the final check, and keep the original audio. A transcript is a working document, not a replacement for the recording it came from.

The bottom line

Nemotron 3.5 Transcribe is not a flashy release, and that is the point. Transcription is infrastructure. When it works, nobody notices. When it fails, meetings become unsearchable, notes become unreliable, and compliance teams get nervous. A free, open-weight model that handles 40 languages, long recordings, and eight speakers gives smaller teams a serious option that used to require an enterprise contract. Try it on your messiest real audio before you decide whether it earns a place in your stack.