Free · Runs in your browser

Two people talking at once? Untangle them.

The best moment of a recording is usually the messiest one. Someone laughs over the punchline, answers over the question, and your transcript turns it into nonsense. PaperEdits turns it back into two clean voices.

Made for: podcasts, interviews, panel talks, wedding speeches, street footage

Sound familiar?

“Both talking over each other. And of course that’s the point where both matter.”cutroomfloor_ed

“Every tool I try completely merges them.”twochairspod

“How do I mute someone while someone else is talking?”first_gig_filmer

Here’s the sneaky part: transcription doesn’t fail loudly on crosstalk. It quietly invents one smooth sentence out of two voices. It reads fine. Nobody said it.

What PaperEdits does

Open your video and say “two people are talking over each other”. It scans the whole thing on your computer in seconds.

Every risky word gets a wavy underline, so you know exactly what not to trust.

Click Untangle. About ten seconds later the transcript splits into Speaker A and Speaker B, side by side, with players to hear each voice alone.

Free to startA free account, and no upload for the in-browser version. Our servers are faster when you want them.

Nearly word-perfectOn our public test recordings, both voices come back almost exactly right.

Works on what you haveNo re-recording, no separate mic tracks needed. One messy file in, two voices out.

Honest about limitsTwo voices at a time, and echoey rooms confuse it. We publish the files that prove both.

Questions
Can it fix my podcast where we talk over each other?

Yes, that’s the exact case. Detect finds the crosstalk, Untangle splits it, and you hear each voice alone before you decide anything.

Does my audio get uploaded?

Not on the free path. Everything runs on your computer. Signing in sends just the overlapping seconds to our servers to do the same job faster.

How many voices can it handle?

Two at a time. Three people at once still gives you two streams, rougher. We’d rather say that here than have you find out mid edit.

For the technically curious

Detection: pyannote segmentation-3.0, 1.5 MB. Separation: ConvTasNet, 20 MB, running as WebAssembly at 3.6x realtime. Word error on our fully overlapped test falls from 72–76% blended to 8%/3% split. The test files are public.

Try it right now