23 September 2026
Guya 4.1: cleaner cuts where it used to be hardest
A new Pleofon update is out for Mac and Windows. The biggest change is under the hood: a new cutting model, Guya CutNet 4.1. Alongside it comes a fix that moved every single cut to where it belongs, plus a few things that simply make editing faster.
Guya 4.1: a model that stopped being afraid of noise
Guya is the part of Pleofon that listens to your recording and decides what stays and what goes: slips of the tongue, repeated takes, hesitations, dead air. On a clean recording from a good microphone it did great. Trouble started when something hummed steadily in the background: wind on a field recording, a car engine, a computer fan, the wash of a big reverberant room.
Why? A model learns from examples, and most of the recordings it learned from were clean. It had simply never met noise.
We couldn't fix that by collecting more noisy recordings. Every one of them would have to be edited by hand first, so the model knew what a correct cut looks like. So we did something else: we built a raw footage generator.
Here is how it works:
- We take finished, edited material where we know exactly what the editor removed and why.
- We rebuild new "raw" recordings from it and splice in real mistakes taken from elsewhere: repeated sentences, abandoned takes, overlong pauses, asides.
- We add what the model had never heard before: noise of many kinds (from a low rumble to a high hiss), mains hum, slight shifts in pace and pitch.
Every one of these examples comes with the answer built in, because we know where we put each mistake. That let the model practise thousands of hard situations that nobody had to cut by hand first. Only then did we fine-tune it on real projects.
The result: on clean recordings Guya 4.1 cuts just as well as the previous version. On recordings with background noise it leaves noticeably less debris behind. On our noisy test recordings that meant between a quarter and more than a third less, depending on the recording.
Silence inside noise
Along the way we found a second, less obvious problem. Pleofon looked for pauses in one simple way: wherever it is quieter than the rest of the recording. With constant background noise it is never truly quiet, so the app could treat half an hour of material as one long sentence. Digest mode, which picks the most important parts, then had nothing to choose from.
Now Pleofon first measures how loud the background itself is in each part of the recording, and looks for pauses relative to that. If the noise changes, say a car pulls away and stops again, the app keeps up. Clean recordings behave exactly as before, and on noisy ones pauses are pauses again.
Every cut, 12.5 ms early
While looking into why cut endings landed where they did, we found a bug that had been sitting in the app for a while. The model rates sound in short windows and describes each window by its centre. The code that then placed the cut read the same rating as if it described the start of the window. The difference was 12.5 milliseconds, on every cut, always in the same direction.
It sounds tiny, but you can hear it. Clips started with a small run-up before the word and ended a touch too early, clipping the tail. With the fix, cuts are noticeably tighter and cleaner. Honestly, the difference surprised us, especially because even before this mishap being noticed - the cuts sounded great already!
Smaller things you'll notice right away
Cuts land on the start of a word. When you drag a clip edge a long way, Pleofon snaps it to where speech actually begins, not somewhere in the middle of a pause. The blade snaps to word starts too.
"Uhh" and "umm" select themselves during cutting. After a cut, Pleofon usually selects the quieter side. But if one side holds a drawn-out, steady sound, that one gets selected, even when it is louder. It's usually the part you want gone.
Pauses without clicks. A pause you insert by hand no longer cuts the sound off with a pop. The voice fades out gently before it and comes back gently after it, both in playback and in the exported file.
Time left. The progress bar now shows how long analysis or project loading will take.
Video files with AAC audio (MOV, MP4, M4A) now open correctly. Some of them used to fail with an "End of file" error.
On Windows
Windows gets the same model and the same fixes. Updating works differently too: a new version shows up as a button in the app's top bar and installs in one click. Pleofon checks for updates on its own at start and every few hours, but never installs anything without your click. The update brings the new cutting model without downloading the large speech model again.