Nvidia released an open transcription model that turns recorded audio into text for free, and you can run it on your own GPU. That makes it a practical building block for captioning video, cutting long recordings into clips, and keeping a searchable transcript of everything you publish. The source: TechCrunch reported the release on October 28, 2026 (techcrunch.com).
raw recording
|
v
+---------------------------+
| Nvidia open transcription | local GPU, no per-minute fee
+---------------------------+
|
+--> transcript.txt ----> clip picks + captions
|
+--> burned-in captions (ffmpeg, local)
|
+--> post text, titles, descriptions
|
v
Creator OS --> 11 platforms + Skool, one API key
|
v
post analytics + watch time, read back by your agent
What Is Nvidia’s Open Transcription Model?
It is a speech-to-text model that Nvidia published openly, so you download the weights and run inference yourself instead of sending audio to a hosted service.
Because the model runs on hardware you control, the audio never leaves your machine. That matters for client calls, unreleased product demos and anything recorded under an NDA.
The practical output is a text file with timestamps. Everything downstream reads from that file: caption timing, clip selection, title ideas, show notes, and a searchable archive of what you have said across hundreds of videos.
One caution. The TechCrunch coverage is the source for this release, and it does not list the model’s size, languages or licence terms. Check Nvidia’s own model card before you build a commercial workflow on it.
Is the Nvidia Transcription Model Completely Free?
The software has no per-minute invoice. Your real cost is the GPU time and the electricity to run it.
That is the honest version of “free.” A hosted speech-to-text service bills you per minute of audio. A local model bills you once, in hardware, and then costs whatever your machine draws while it works.
| Cost | Local Nvidia model | Hosted speech-to-text |
|---|---|---|
| Per minute of audio | Nothing | Billed by the provider |
| Hardware | Your GPU | Provider’s servers |
| Audio leaves your machine | No | Yes |
| Setup work | Yours | Account and key |
| Runs offline | Yes | No |
If you already own a GPU for video editing, the transcription step adds no new hardware. That is the case where “free” is literally true.
How to Run Nvidia Transcription Locally Step by Step
The shape of the workflow is always the same: install the runtime, download the weights, point it at a file, write the transcript next to the video.
Because I only have the TechCrunch report, I will not invent package names or command flags for the model itself. Use Nvidia’s own quickstart. The structure below is what you are aiming for.
- Confirm your GPU driver and CUDA runtime are current, since the model runs on the card.
- Create a Python environment for the project so the model’s dependencies do not collide with your editor.
- Pull the weights from Nvidia’s model page and note the licence terms.
- Run the model against a short test clip first, not your two hour podcast.
- Save the output as plain text with timestamps next to the source file.
A file layout that survives contact with real projects:
project/
raw/episode-014.mp4
transcripts/episode-014.txt
captions/episode-014.srt
clips/
Once transcription works on one file, wrap it in a loop over a folder. That is the moment it stops being a novelty and becomes part of your pipeline.
How to Turn Long Videos Into Clips With Free Transcription
Read the transcript first and pick clips by what was said, not by scrubbing a timeline. The text is a faster map of a 90 minute recording than the waveform is.
A prompt you can hand to any MCP capable agent once the transcript exists:
Here is the transcript of a 90 minute interview with timestamps.
Pick 6 moments that stand alone without setup.
For each one, give me the in and out timestamp, the hook line
for the first 3 seconds, and a caption under 120 characters.
Avoid anything that needs the previous 30 seconds of context.
The agent returns six time ranges. You cut them, caption them, and publish. The transcript did the hard part.
If you want the editing half to stay local too, read an AI video editor built on ffmpeg and Remotion for a walk through what that style of local workflow looks like. For choosing a card before you commit, see picking a local video model by GPU VRAM. Both are posts on this blog, not tools Nvidia ships.
What Open Source Video Editor Pairs With Auto Captions?
Pair the transcript with a local editor that renders captions from a file, so the caption step never becomes a manual typing job.
Creator OS ships OPEN VIDEO EDIT, a set of open source Claude Code skills that turn a raw talking head clip into an animated short. Captions, logos, b roll cards and counters are part of the render, and it runs locally with Python and ffmpeg.
The skills include split animated talking head, longform animated talking head, fireship style edit, fireship style short, split animated short and reverse engineer. The fireship style pair is the one to look at if you want fast cuts with captions on screen.
Put the two halves together and the loop is: Nvidia model writes the transcript, OPEN VIDEO EDIT burns the captions and animations, you review, you publish. No per minute charges anywhere in that chain.
Kevin walks through the tooling on the KevBuildsApps YouTube channel, and the Creator OS launch video shows the publishing side in motion.
How Transcription Data Feeds Your Post Analytics
A transcript is not just a caption source. It is a record of what you actually said, and that record can be compared against how each post performed.
Here is a concrete loop. Transcribe ten videos. Tag each clip with the topic it covers. Publish through Creator OS. Then pull the numbers back per post.
creatoros analytics:posts --from 2026-10-01 --to 2026-10-31 \
--sortBy engagement --limit 20
For video, the useful fields are on the post itself. get_post_analytics returns Reel watch stats per post: reelsSkipRate (the share of views that swiped away in the first 3 seconds), igReelsAvgWatchTime and igReelsVideoViewTotalTime in milliseconds. Meta reports these for Reels only.
Now the transcript earns its keep. If the clips built on hook style A show a lower skip rate than hook style B, you have evidence, not a hunch. Your agent reads the numbers and your own instructions decide what changes in the next video.
The same idea works at the account level with get_best_time_to_post and get_daily_metrics. Posting time plus transcript topic plus watch time is a real dataset, and it is all reachable from one key.
Nvidia Transcription vs Paid Speech-to-Text: What’s the Difference?
The difference is who holds the audio and how you pay. Both produce text. They differ on privacy, on cost shape, and on how much setup you accept.
A paid hosted service is a dashboard or API you sign into. It is fast to start and someone else handles the GPUs. Your audio is uploaded to their servers, and the bill scales with minutes.
The local open model is the reverse. Nothing is uploaded. The bill does not scale with minutes. You own the install, the driver updates and the failed runs.
The same distinction shows up in publishing. A social media scheduling dashboard asks you to click through a calendar. Creator OS ships a hosted MCP server and a CLI so an AI agent can run your accounts, which means the transcript, the clip and the publish call can live in one script.
creatoros posts:create --text "Hook line from the transcript" \
--platforms instagram,tiktok,youtube \
--media ./clips/ep14-hook-a.mp4 \
--scheduledAt 2026-11-03T09:00:00Z --timezone America/Toronto
That command is real. The platforms are real. What the agent decides to put in the text is where the transcript pays off.
FAQ
Is Nvidia’s open transcription model really free to use commercially?
The TechCrunch report does not state the licence terms, so check Nvidia’s model card before commercial use. The software has no per minute fee, but your GPU time and electricity are real costs.
What GPU do I need to run Nvidia transcription locally?
The release coverage does not list hardware requirements, so confirm them on Nvidia’s model page. If you already run local video models for editing, you likely have a card in the right class. Our post on picking a local video model by VRAM walks through choosing one by memory size.
How accurate is Nvidia’s model for multilingual video captions?
I cannot give you an accuracy figure because the source report does not include one. Test it yourself on a two minute clip with accents, crosstalk and background music, since those are where caption errors hurt most. Always read captions before publishing.
Can I auto-caption videos with a free open source video editor?
Yes. OPEN VIDEO EDIT renders captions and animated overlays locally with Python and ffmpeg, and it accepts a talking head clip plus its transcript. It is open source and installs as Claude Code skills.
How do I turn long recordings into short clips with transcription?
Transcribe the full recording, then ask an agent to pick self contained moments with in and out timestamps and a hook line for each. Cut those ranges, caption them, and schedule them through Creator OS. The transcript makes the selection step fast and repeatable.
Get started
Local transcription gets you a transcript for free. Publishing that transcript across every platform still needs a place to send it.
Creator OS publishes, schedules, replies and reads analytics across 11 platforms plus Skool, from the web app, the iOS app, a REST API, the CLI and a hosted MCP server for Claude, ChatGPT, Claude Code, Cursor and other MCP capable apps. Sign up and connect your accounts, then start with creatoros auth:check and creatoros accounts:list to confirm the wiring. The CLI docs cover the rest of the commands, and the docs cover the API base URL and request shapes.
Creator OS plans start at $14.99/month for 11 platforms plus Skool, with the API, the MCP server and the CLI included.
If you want to see the whole loop running, transcript to caption to published post, watch the walkthroughs on the KevBuildsApps YouTube channel.
Get started for $0
Free trial for new subscribers. Cancel anytime.