Zoom recordings → text, free.

Best use: your class or meeting recording on Zoom has a transcript, but the host turned off downloading. Copying the transcript out of Zoom's side panel usually breaks: lines go missing, timestamps and names get jumbled, or it only copies what's on screen.

Paste the link here instead. If Zoom has a transcript, you get it in seconds as a clean file:

• .txt for notes and searching, .srt or .vtt for captions in video players
• With timestamps, ready to copy in one click

No transcript on Zoom? The site makes one itself. It listens to the recording on your own computer using a free, open-source speech model. This takes longer (see How to make it faster), but it's free and private. No account, and nothing is stored or sent to an AI service.

1. Look for Zoom's own transcript first. It's instant. If you're the host, turn it on under Zoom → Settings → Recording → Create audio transcript.
2. Use a laptop or desktop with Chrome or Edge, plugged in, with power mode set to Best performance.
3. Make Chrome use your dedicated graphics card (if you have one): Windows Settings → System → Display → Graphics → Chrome → High performance, then restart Chrome. This is often the biggest speed-up.
4. Pick the spoken language instead of Auto-detect.
5. Need a rough draft? Choose Balanced or Fast. If it's going to take long, the site also offers to switch partway through.
6. The first run is the slowest because the speech model downloads once and is saved. Don't clear site data or use Incognito, or it downloads again.

Rough guide: with a good graphics card, an hour of speech takes a few minutes to ~20 minutes. On basic laptop graphics it can take about as long as the recording. Silence is skipped automatically.

• Not 100% accurate when the site transcribes the audio itself. Names, numbers, jargon, accents, crosstalk and poor microphones cause mistakes, so double-check anything important. (Zoom's own transcript has its own mistakes too.)
• No speaker names in the site's own transcription. Zoom's transcript includes them when available.
• Sign-in-only links don't work (Zoom login or school SSO). Download the file from Zoom and upload it instead.
• Long recordings take a while on computers without a dedicated graphics card, about as long as the recording itself.
• Phones and old computers are too slow for long recordings.
• Keep the tab open until it finishes. Closing or reloading the tab stops it, but you can download what's done so far first.
• Link mode depends on Zoom's website. If Zoom changes it, use Upload a file, which always works. See the status page.
• Only transcribe recordings you're allowed to access.

This project runs on free hosting with no servers to pay for. That's why it's free and private, and also why some things are out of reach. Here are ideas for the next person, roughly from most to least impactful.

1. Faster transcription in the browser. The biggest pain point. The site processes one 30-second piece at a time, and the browser adds overhead for every word the model produces. Running several pieces at once on the graphics card (batching), or newer browser-optimized speech models, could give a 1.5–3× speed-up. Measure on typical student laptops (built-in Intel/AMD graphics), not just gaming PCs.
2. An optional "fast lane" server. The same model runs several times faster as a native program than in a browser. A volunteer-run computer with a graphics card (a lab PC, a club server) could take a queue of links and email results. That needs someone to host it, a privacy policy, and abuse limits, which is why it isn't here now.
3. Sign-in-only recordings. A small browser extension could run inside the student's own logged-in Zoom tab and grab the transcript or audio there, which the website can't do. It would cover most school (SSO) recordings.
4. Pick up where you left off. Save progress in the browser so a closed tab or a sleeping laptop can resume instead of starting over.
5. Speaker labels. Open-source speaker-separation models exist but are heavy for a browser. Another option is to reuse names from Zoom's transcript when it exists.
6. Smarter silence detection. The current check is based on loudness and errs on the side of keeping audio. A small dedicated voice-detection model would skip more non-speech (music, room noise) without losing quiet speakers.
7. Keeping link mode alive. Zoom doesn't officially support fetching recordings this way, so it can break when Zoom updates its site. The daily check emails the maintainer. The same logic is maintained in the open-source tool yt-dlp, which is a good reference when fixing it.
8. Phone support. Phones are too slow for long recordings. A "send this to my laptop" hand-off, or small models for short clips, would help.
9. Stay true to the idea: free, private, and no AI APIs. Features like summaries or study notes would normally need a paid AI service. If added, make them optional and clearly labeled.