AIEXTRACASHALPHA中文
HOME / NEMOTRON

NVIDIA Nemotron 3.5 Transcribe Guide: Free AI That Turns Recordings Into Speaker-Labeled Text in 40 Languages

2026-10-28Last verified 2026-10-28

NVIDIA Nemotron 3.5 Transcribe: Free AI That Turns Recordings Into Speaker-Labeled Text in 40 Languages

In one sentence: imagine hiring a stenographer who never sleeps, understands 40 languages, writes down who said what in a meeting, and works for free — as long as you give them a decent NVIDIA graphics card to sit at.

Two terms, plainly: open-weight means NVIDIA published the model’s actual files for free — anyone can download and run them, including for paid client work. Transcription just means converting spoken audio into written text. Speaker attribution (also called diarization) means the transcript labels each turn — “Speaker 1:”, “Speaker 2:” — instead of one undifferentiated wall of words.

Released October 28, 2026, per NVIDIA’s official blog post that day. One honest gate before anything else: this is not a browser tool. There is no free web demo we could verify on launch day, and running it needs an NVIDIA graphics card plus a few terminal commands. If you don’t have one, skip to the “No NVIDIA GPU?” section at the end — the money play still works, just with a different tool.

Why it matters to you

  • A 1-hour meeting becomes searchable notes. Instead of re-listening to the whole call, you search the transcript for the decision.
  • One recording, four outputs. The same file can come back as a plain transcript, a translation into English, a speaker-labeled script, or answers to questions like “What decisions were made in this meeting?” — you steer it with plain-language instructions, per NVIDIA.
  • Podcasts and videos get subtitles and show notes. Creators sit on hours of un-captioned audio; this does the heavy lifting for $0 in model cost.
  • Private by design. Because it runs on your own machine, the audio never uploads to a vendor’s cloud. What this means for you: a client’s confidential meeting stays on your computer.

The numbers, with vendor claims labeled

  • 40 languages transcribed; 31 translated to or from English; speaker attribution in 13 languages. Recordings up to 1 hour long, up to 8 speakers labeled.
  • 30 billion parameters total, about 4 billion active at a time — think of it as a big team where only a small crew works on each word. That is why NVIDIA says it runs on a single GPU, from a Jetson AGX Thor or DGX Spark machine up to a GeForce RTX 5090 or a data-center Blackwell card.
  • Accuracy claims are NVIDIA’s, not independently verified: NVIDIA says Transcribe ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English (5.33% average word error rate — roughly 5 words wrong in every 100) and ties for first among open-weight models on Artificial Analysis at 2.8%. What this means for you: on public test sets, NVIDIA’s own numbers put it at the top of the free-model pack — but treat any leaderboard claim as a vendor claim until independent testers reproduce it.
  • Recorded audio only, not live streams. NVIDIA and its launch coverage agree on this limit.

Price: $0 for the model

The weights are free to download under the OpenMDW-1.1 license, which allows commercial use. What this means for you: you can legally charge clients for transcription work done with it and keep 100%.

What you do pay for: the hardware. A capable NVIDIA graphics card (or time on a rented GPU machine) and electricity. Hosted pay-as-you-go endpoints were mentioned by NVIDIA’s partners at launch, but we could not confirm a hosted price on launch day, so we quote none — if a provider’s price page is your deciding factor, check it directly before committing.

Setup in 5 steps (you need an NVIDIA GPU)

Honest upfront: this involves a terminal and a download measured in gigabytes. No coding knowledge needed — just copy commands from the project’s own instructions — but it is heavier than a Mac app install.

Step 1. Confirm you have an NVIDIA graphics card. On Windows, open Task Manager → Performance → GPU and check the name says NVIDIA. No NVIDIA card? Jump to “No NVIDIA GPU?” below.

Step 2. Go to NVIDIA’s official announcement and follow its links to the model files on Hugging Face (the blog post links the model; the research page lists the checkpoints: BF16, FP8, NVFP4, and a GGUF build for local runners).

Step 3. Pick your runner. NVIDIA names two: vLLM (for serving the model) or NeMo-Speech.cpp (for running it locally). The GGUF checkpoint is the one built for local, on-device running.

Step 4. Download the checkpoint and run the project’s own example transcribe command from its README — copy it exactly rather than typing it from memory. If the command fails, the two usual culprits are a wrong file path and a GPU running out of memory; try the smaller quantized build (FP8 or NVFP4) first.

Step 5. Try your first real prompt. NVIDIA’s own examples, copied from the announcement:

  • “Transcribe this recording and label each speaker.”
  • “Translate this audio into English.”
  • “List what decisions were made in this meeting.”

What you should see: a text transcript with speaker labels, a translation, or a short decision list — depending on which instruction you gave. Spot-check names and speaker changes against the recording: NVIDIA itself warns that speech models still make mistakes on names and speaker switches.

The money play: the meeting-notes transcription gig

  • Who pays: podcasters, YouTubers, coaches, consultants, and small teams sitting on hours of recorded calls with no notes.
  • What you sell: “I turn your recordings into clean, speaker-labeled notes with decisions and action items, 48-hour turnaround.”
  • What to charge: set your own test price — start low for your first 3 clients (a small per-audio-hour rate), then raise it once you have samples to show. This is a suggestion to test the market, not a market rate. Your cost per job is $0 in model fees, so the margin is your time.
  • 7-day validation: Days 1–2, set up and transcribe 3 public recordings (a podcast episode, a public meeting, a lecture) as portfolio samples. Days 3–4, message 20 podcasters or creators offering one free 5-minute sample transcript. Days 5–7, convert to paid — the goal is 1 paying client or 5 warm leads. If nobody replies after 20 genuine outreaches, drop it: the setup cost you nothing but time.

Pitfalls and limits

  • No live transcription. Record first, transcribe second. A client who needs real-time captions needs a different tool.
  • GPU-gated. A Mac without an NVIDIA card is not NVIDIA’s supported path for this model; the GGUF build plus NeMo-Speech.cpp is the local route, and it still assumes NVIDIA hardware in NVIDIA’s own docs.
  • Names and speaker switches need spot-checks. NVIDIA says so itself. Never send a client a transcript you haven’t skimmed for misheard names.
  • No verified hosted price yet. If you plan to run this for clients on rented GPUs, price that before you quote them — don’t assume a per-minute rate.
  • Language support differs by task. Transcription covers 40 languages, but translation and speaker attribution cover fewer (31 and 13). Check the task-by-language matrix on the model card before promising a client a language.

No NVIDIA GPU? Do this instead

Use the same gig play with a tool that fits your machine: our Granite Speech 5.0 guide runs free on an Apple Silicon Mac, and Koi is a $59 Mac app that does dictation and meeting notes with no terminal at all. And if you want the talking side of the Nemotron family instead of the listening side, see our Nemotron Bolt E4B guide — a free conversational model that runs on almost any GPU.

Sources

Don't just read it — run the first job this week

Subscribe and get a 7-day validation checklist.

Subscribe free