RELEASE NOTES
When two people are in frame the app finds the speaker. Not a separate setting
— it is automatic inside both « Smart (face) » and
« Face tracking ». If more than one face is visible, the app compares
sound with picture: the person whose mouth movement lines up with the audio envelope is the
one speaking, and the frame goes to them. The frame does not run off because someone moved
their mouth for nothing. When the speaker changes, it cuts there. On single-person scenes
this step never runs (so you never wait for nothing); the first time it is needed the
vision engine downloads once (~200 MB).
Face detection is far better in dark scenes. The search ran on a small image with
no contrast correction, and in dark studio footage it found no faces at all.
Measured: on the same three frames the old settings found 0 faces; the new ones find them
all.
The clip is now SPLIT at a camera cut. When the source video switches between two
speakers, the app finds that moment, cuts the clip in two right there and centres each
piece on its own face. No drifting, no animation — because the picture is already
cutting, a framing change there goes unnoticed. Measured on a real project: in a
four-minute video 23 of 23 camera cuts were found.
Face framing is measured far more often. Previously at most 12 samples were taken
per clip; with « no cutting » selected that meant one measurement
every 21.5 seconds on a 237-second video — most cuts were invisible and the
framing settled on the average of long stretches. It is now measured about every 0.45
seconds. Face detection also runs on a larger image with contrast correction (because the
framing magnifies the picture 3.16×, a 1-pixel error became 13 pixels on screen).
« Smart (face) » framing is also set per scene. Previously
three samples were taken across the clip and their average was used as a single crop
position; once a cut came in between, that average landed exactly halfway between the two
speakers — that is, on nobody.
← Back to the documentation
What changed, point by point
Something wrong in the new version?
Inside the app, choose Updates → Roll back to the previous version. It downloads the installer, verifies it and installs it for you. Your downloads and settings stay exactly as they are.
You can also download any version manually from the list below.
What recent releases gave you
- One-file setup: run the exe, « next, next, install » — no extra downloads, no manual steps.
- The same reliability on every machine: the bugs around long file paths, non-Latin user names and folders with spaces are closed.
- Processing is now in plain sight: minimise it, send it to the tray, and see the time elapsed and remaining.
- Nothing fails quietly: if a tool is missing from the install, it is now named.
- Fixed — three engines could not be installed on a new computer. The makers of the automatic subtitle, audio cleanup and offline translation engines changed how they publish their files, and the installer could no longer find them. It now recognises the new formats too.
- Fixed — « Video removed » on a direct video link. With a quality selected, a direct .mp4 link was not downloaded even though the video was there. The closest available format now downloads, and if something really is wrong the error names the actual reason.
- The installer speaks 13 languages. Component names, shortcuts and status messages now appear in the chosen language; if your system language is on the list, setup starts straight in it.
- 🎬 Automatic subtitles no longer make things up over music. Speech recognition kept rewriting earlier sentences in musical passages (357 words for 97 spoken words in one 2-minute stretch). Measured against the official subtitles of Tears of Steel, the error rate fell from 146% to 21%, fewer lines are skipped and subtitles are ready sooner. The censor and dubbing use this cleaner text as well.
- 🌍 Better translation on a powerful graphics card. On a computer with an NVIDIA card of 20 GB or more, offline translation now runs a larger model (12B) on the graphics card. Measured against the film's official translations: more accurate in all 9 languages (most clearly in Spanish) and about 7 times faster. The model downloads once (about 8 GB); if you already installed translation, it arrives by itself in the background the first time you use it. Without such a card, or when it is busy, today's model keeps working.
- Fixed — uninstalling left the translation and smart caption models on disk (up to 5 GB in total).
3.9.425
SEPTEMBER 2026- 🎙 Dubbing — make a video speak another language. Turn on « Dubbing » in Editing & Subtitles and pick a language: an English video speaks Spanish, Turkish or any of the 13 supported languages. The original speech is removed; music and effects stay. Each speaker gets their own voice (female and male kept apart), and subtitles, if on, show the dubbed text.
- « Voice-over » type: the original sound isn't removed completely — it stays underneath, turned down, like a documentary narration.
- 🔊 Read text aloud. Text overlays can now be read out: 6 voices, 5 tones, the language is detected or chosen by hand. The video's sound dips while reading, and the text can stay on screen until the reading ends.
- ✨ In Super Edit too. A « Dubbing & Read aloud » group in AI auto-edit, and a « Voice it » button in the Inspector for the selected text. The dip is now audible in the preview as well.
- 🚫 Slang and swearing censor, now for downloads too. A box in Editing & Subtitles mutes or bleeps swear words and masks them in the subtitles — even when subtitles are off. The word list got tougher: spellings without accents, with digits or split into pieces are caught too.
- ✨ Smart subtitles. « Automatic emoji and emphasis word » enlarges the key word of each sentence and drops an emoji in the right place; « Opening title » puts a short title, taken from the speech, behind the person in the first seconds. Runs on your computer, no key needed.
- Skipped lines are listened to again. Speech recognition sometimes passed over repeated chorus lines without writing them; those gaps are now listened to separately and added to the subtitles and the censor.
- Optional install: the voice engine is large (6-9 GB). Choose « Read-aloud and dubbing » in the installer, or skip it and the program installs it the first time you use the feature. Dubbed files carry an « AI voice » label.
- ⚡ Reading check twice as fast. Every reading is played back to confirm it was spoken correctly, and read again if not; that check took most of the time. It now uses more of the processor's cores and finishes in half the time — same result.
- Fixed — sound ran ahead in videos with a cover. When a cover image was added at the start, the audio started early by the cover's length. They now start together.
- Fixed — speaker separation didn't work at all on some installs. The first letter of the model path was cut off when it was read, and the feature quietly switched itself off.
- Fixed — music cut out in Super Edit exports. Background music went silent as soon as a narration or reading ended.
- Fixed: « Human voice only » silence cutting now really detects speech · « Stop » works during speaker separation · the translation engine shuts down 10 minutes after the job and frees memory · the spoken-language list now covers all 13 languages.
3.9.424
SEPTEMBER 2026- 🔑 The « HuggingFace Token » setting is gone from Preferences. Speaker separation has run on the built-in voice-fingerprint engine since 3.9.420 and needs no key; the field had simply been left behind in Preferences → AI Settings.
- The false warning is gone too. Ticking « Separate speakers » still produced « no HuggingFace key — subtitles will come out normally ». The feature works fine without a key, and the warning was talking people out of using it.
- The Gemini, OpenAI and Claude keys are still there — those are still used.
3.9.420
SEPTEMBER 2026- 🗣️ Separate and colour the speakers — it actually works now. In podcasts, interviews and multi-person streams each speaker is written in a different colour. The app works out how many people there are and who spoke when by itself — it recognises them from the voice fingerprint and, when faces are visible, confirms it from the picture.
- The old one never worked at all. The checkbox existed, but the engine side depended on a component that was not installed and on a HuggingFace key; ticking the box did nothing and said nothing about why. Neither the key nor that component is needed any more.
- The « minimum / maximum speakers » boxes were removed. Asking you for the number of speakers meant « if you do not know, you get a bad result ». One checkbox is left; everything else is automatic.
- Compatible with every design and presentation. The colour changes only the text itself; the outline, shadow and box keep the colour of the design you picked. It works in all four presentations — sentence, single word, karaoke and stacking. If the karaoke highlight clashes with a speaker colour the highlight turns white on its own — otherwise the highlighted word became invisible.
- The preview shows it too. Tick the box and the preview alternates between two speakers; you see what you will get before downloading.
- On first use a ~25 MB recognition component downloads by itself; you do not have to run anything.
3.9.417
SEPTEMBER 2026- 🔴 Translation could not be selected — fixed. The « Subtitle language (translation) » list stayed empty: the file that builds the language table was not being loaded into the popup. There was no error message either — the list quietly stayed on « No translation ». 10 target languages can now be chosen.
- New: subtitle PRESENTATION. Until now there was only one — « print the sentence ». There are now four presentations, independent of the design: Sentence · Single word (one word at a time, large — gaming/TikTok) · Karaoke (the sentence stays, the spoken word grows and changes colour) · Stacking (words are added one by one). Any design can be combined with any presentation.
- New: subtitle size. There was no setting before; every design came out at its own fixed point size. There are now five steps from 75% to 145%, and on boxed designs the box grows with the text.
- Four new designs. All ten of the old ones came from the « thin outline + shadow » family. The new ones are different families: Gamer (very heavy outline, ALL CAPS), Highlight Box (a solid coloured box behind the spoken word), Heavy Outline (no shadow, the most readable), Lower Band (small and calm — documentary).
- The preview shows all of it. Size and presentation feed into the live preview too — you see what you will get before downloading.
3.9.413
SEPTEMBER 2026- Subtitles no longer disappear at a cut. When the automatic edit split the video into pieces, a subtitle that straddled a cut showed up in only one piece; on the other side the screen stayed blank while speech continued, and the next line then looked like it « arrived late ». Measured on a real project: 11 of 14 subtitle gaps began exactly at a cut point, and 59 words were spoken without subtitles in those gaps (the longest 2.7 seconds). A subtitle now appears in every piece it touches, and the text is divided in proportion to how long each piece is heard. Coverage measured on the same edit: from 78% to 100%.
- Instruction text no longer leaks into the subtitles. While repairing reading speed, the AI sometimes returned the rule it had been given instead of rewriting the sentence, and a line like « Within 16 characters at most. » appeared on screen. A shortened line is now rejected if its content does not match the original — both in Super Edit and on the download screen.
- Half-lines left stranded on screen are merged. One-word fragments like « instead, » or « and this too. » sat on screen for a second and a half. They now join the neighbouring line — as long as that does not break the reading speed or the fit on screen. In the same video, all nine orphaned fragments were closed.
- Reading-speed repair. When a translation grows too long for one line, the app has the sentence rewritten shorter (in professional subtitling this is called condensation). Measured on a real project: the share of lines over the comfortable reading limit fell from 80% to 30%, and the median reading speed from 25.8 to 15.2 characters per second.
- Subtitles can now be translated — on the main download screen too. Even when the video is spoken in English, you can get the subtitles in another language. This used to exist only inside Super Edit; the « Language » field on the download screen selected the spoken language, and because of its name most people took it for the target language. There are now two separate fields: Spoken language and Subtitle language.
- Subtitles no longer arrive late. Translation is now done on the whole sentence and then fitted back onto the spoken word timings. Previously 3-5 word fragments were translated one by one; because the verb goes last in Turkish the meaning arrived late and lines made no sense on their own (things like « these variants. However these two »). Measured: in the same video the share of lines over the reading-speed limit fell from 25% to 0%.
- Translation without a key. The app now has its own translation model: no internet account and no cost. It downloads once on first use (about 2.3 GB) and the app handles that itself. Translation in the download flow always uses this model — your API key is never handed to the processing engine. In Super Edit you can pick Gemini or Claude if you prefer; the Test button tries the engine you chose on a real sentence.
- Each language gets its own reading speed. Subtitles are now broken into lines according to the target language's reading speed. That speed varies a lot — 17 characters per second in Turkish, 20 in Arabic, only 4 in Japanese — and the app previously knew nothing about the difference.
- Russian, Arabic and Chinese subtitles no longer vanish. Inside the engine there was a gate that checked whether a subtitle was « meaningful », and it only counted Latin letters. Measured: a 285-line Russian subtitle could not get through that gate and was silently thrown away with « is there no human speech in the video? ». Even in Turkish, everyday words carrying accented letters were not counted. Chinese and Japanese were worse still: the subtitle file came out completely empty and the app still reported « success ».
- Subtitles no longer spill outside the frame. The engine never wrapped long lines. Measured: in a real Turkish translation 74% of lines were cut off at the edge of the screen, and 86% in Russian. Lines are now broken in the right places.
- A failed translation no longer reports « success ». Previously, even when translation failed outright, a green « ✓ 143 subtitles added » appeared; you picked Turkish, got English subtitles, and saw no error at all. A half-finished translation was silent too. Both are now reported, with the reason.
- Small but annoying: the subtitle section told you to run a file that does not exist. The app installs the engine it needs by itself; that wrong instruction is gone.
3.9.411
SEPTEMBER 2026- The Freeform Composition editor now looks like the rest of the app. This screen used a green accent; Super Edit is blue and Multi-Panel was purple until last release — the app was carrying three separate colour worlds at once. They are all the same now. The icons are the app's own line icons rather than emoji (this screen had 114 emoji). A version badge was added — there had never been one.
- Pieces snap to one another. While dragging, they now snap not only to the edges of the frame but to the edges and centre lines of other pieces, with a guide line showing which line they snapped to. Hold Shift for fine adjustment and snapping turns off.
- Align and distribute tools. Align left, centre, right, top, middle and bottom, plus distribute with equal spacing. If there is a selection it applies to the selected pieces, otherwise to every visible piece.
- Multiple selection. Shift+click selects several pieces and aligns them all at once. On a screen that allows twenty pieces, they could only be selected one at a time.
- Piece lock. Once a piece is placed you can lock it; a locked piece cannot be dragged, moved with the keyboard or affected by alignment. The background piece used to shift accidentally on every drag.
- Keyboard navigation. Tab cycles through the pieces.
- The preview cover sent to the popup is no longer incomplete. The app counted a fixed amount of time instead of waiting for the frames, so on a twenty-piece composition the cover could be saved blank. It now waits until the frames are genuinely ready.
3.9.408
SEPTEMBER 2026- Saved projects open again. Project files saved in Super Edit since 26 August said «project content is corrupt» when opened. Nothing inside the files had been lost — the way they were written had simply drifted from the way they were read. Your existing files open as they are with no conversion needed; across six test files every layer, clip and keyframe came back.
- Your project presets open too. For the same reason, saved project presets would not load. The presets were never lost; they were sitting right where they were. Item and animation presets were unaffected.
- «Continue where you left off» is back. The autosave that restores your work when the editor window is closed by accident was silently disabled for the same reason.
- A file that will not open no longer wipes your work. If you picked a corrupt file or one belonging to another program, the app emptied your current edit first and reported the error afterwards. It now checks the file first; if it cannot be read, your edit is left untouched and you are told plainly what happened.
- Old Pitch / Echo settings are preserved. Opening a project saved before 31 August silently discarded those two audio settings; they are now carried into the audio effect chain on load.
3.9.407
SEPTEMBER 2026- Beats now land exactly where they belong. The rhythm markers that were found came about 40 milliseconds ahead of the real beat in the audio — that was the «it is rushing» you could hear. Measured and fixed: on the kick channel the deviation went from 41 ms to zero, and the share of beats off by more than 30 ms fell from 88% to 31%. Which beats are found did not change — only where they land.
- The channel that vanished on fast tracks is fixed. An instrument with more than 260 beats per minute (a hi-hat in genres like hardtekk or drum&bass) was silently dropped from the list. The limit was raised; on a test track a treble channel at 287 beats per minute came back.
- The app no longer looks «frozen» while analysing. The buttons lock, the one you pressed shows a working indicator, and the elapsed time counts up second by second. The first analysis takes a few minutes (the audio is split into instruments); later searches on the same track come back in seconds.
- It does not miss beats in quiet passages. If the intro or a breakdown was low in level, its beats never showed up at all. The app now opens its ear in those stretches; on the test track the beats in the first seconds were caught and everything else stayed exactly the same.
- The panel got simpler. A single Find rhythms button pulls out all four instruments at once, with individual search buttons underneath. Each rhythm gets Beep (check it by ear), Mark (put it on the ruler), Cut (cut on the beats) and Remove. The measurement line is readable now: how many beats, how often, how many BPM.
- Marker Tools is its own section. Selecting, copying, pasting, thinning, shifting and rubber-band selecting markers all live under Mode ▶ Marker Tools.
3.9.379
AUGUST 2026- Face tracking was rewritten — it does not lose the face any more. The old version could only recognise faces looking straight ahead: turn to the side, tilt your head or move fast and the frame lost you. Measured on a real dance video: a face could be found in only 23% of frames. It now works in three stages — frontal face, wide-angle face, and if neither is found, head position from body pose. Even with the person's back turned, the frame holds the head. If no face is visible at all it follows the person themselves and never freezes.
- And it is no longer late. The framing is measured eight times a second instead of twice, and the motion is computed both forwards and backwards so the lag comes out at zero. If the face vanishes for a moment the frame does not freeze and then jump — it glides between the two points. With several people in shot it no longer bounces between them either: it follows whoever it started on.
3.9.374
AUGUST 2026- Audio effects arrived — Library › Audio FX. Echo, reverb, pitch (up/down), telephone, radio, robot, chorus, tremolo, vibrato, bass, treble and lo-fi. Select an audio item and click an effect; the effects form an ordered chain and you can rearrange it — the order changes the result. Each effect can be bypassed temporarily so you can hear before and after. Because a clip's sound lands on its own audio item when you add it, these apply to clip audio too.
- You can hear them in the preview now. Earlier versions had two settings called « Pitch » and « Echo », but they did nothing in the preview — they were only applied on export. So you moved a slider, heard nothing, and the sound came out different in the finished video. Every effect is now audible during playback. Pitch/echo settings in your old projects are carried into the new chain on load; nothing is lost.
- The cards draw what they do. Sound cannot have a visual preview, so each card shows what the effect does to the waveform: decaying pulses for echo, a square wave for robot, an undulating envelope for tremolo. Hover over a card and it also tells you when to use it.
3.9.359
AUGUST 2026- Settings with no effect on the output — four of them fixed at once. This release was given over to a single class of bug: you change a setting, you see it in the preview, it has no effect at all on the output, and nothing anywhere tells you.
- The sound of a video you added never reached the output. The interface had a « Volume » slider, but that audio was not mixed in on any path. Measured: the overlay's sound was completely absent from the output; now both are heard together with the main audio not ducked at all. (If you add a file with no audio, the app notices and skips it.)
- The « slide » animation did not slide on an overlay with a time range. Say « appear at 10 seconds, slide in from the left » and the overlay simply appeared at its destination. The reason: the animation time was measured from the start of the video, not from the moment the overlay appears. Bounce and shake were affected by the same thing.
- Opacity, rotation and animation did not work in the Start/End placement. The same settings worked in the normal placement and were visible in the preview — they were just missing from the output. Two separate code paths had drifted apart; both are now fed from the same place.
- The files behind overlays saved into a preset went missing. After closing and reopening the app, applying the preset said « 3 overlays included » but the files did not arrive, and pressing Download stopped the download entirely — waiting did not help either. The files now come from the app's own media vault; if a file really is missing, you are told which one.
3.9.357
AUGUST 2026- The app window is actually used now. The interface was pinned to a 460-pixel column written for a browser popup, and the desktop app showed the same column: in a 900×955 window, 820 pixels sat empty on the right and 302 at the bottom. The shell now fills the window. The most visible result is in the Overlays section: in a wide window the preview opens on the left with the layers and settings on the right — the section's wide layout is genuinely usable for the first time.
- Two windows now see each other instantly. Overlay changes were only carried across when you came back to a window; with both open at once the focus never changed, so one stayed unaware of the other. Measured: nothing arrived for 6 seconds before, now 0.4 seconds. No handover happens mid-drag — the work in your hands is not interrupted.
- The Overlays section is smoother. The entire interface was rescanned every time the layer list was drawn. Only the part that changed is scanned now: across 60 draws, 1125 queries → 160, and nodes scanned 10,914 → 792.
- Silent losses were closed. Deleting on one clip and then moving to another made undo restore the wrong clip's list. « Apply to all clips » deleted the overlays on the targets without asking — it now says how many will go and asks first. And an overlay whose file could not be found was quietly dropped by the engine: it is now named in the log and in the summary line.
- Small but visible fixes. The « No overlays yet » box appeared twice side by side; one was removed. Words ran together in the setting labels (« KEY(CLIP… ») — the space is back.
3.9.348
AUGUST 2026- The Super Edit preview is smooth on heavy projects too. As subtitles, cuts and
transitions piled up on the timeline the preview stuttered, while export was fine. Measured:
drawing was not the culprit — painting a frame with every effect on takes 0.28
milliseconds (the 60 fps budget is 16.7 ms).
The real cause: moving a slider or typing a subtitle made the app delete the whole timeline and rebuild it from scratch. On a 180-item project that takes 52 ms, on 720 items 222 ms, and during that time the browser can do nothing else — which is why the preview froze. The rebuild now happens only when genuinely needed, once, when you release the slider; settings with no visual counterpart on the chip (outline, box colour, background blur) do not rebuild the track at all; and typing a subtitle updates only that clip's label. - The measured result. Frame time across a 30-step slider drag (on projects of 60 / 180 / 360 / 720 items): 25 · 75 · 150 · 261 ms → 3.6 ms in every case. The same while typing a subtitle: 25 · 82 · 143 · 257 ms → 3.6 ms throughout. Editing is now smooth regardless of project size.
- The ruler no longer forces a layout calculation on every draw. The timeline width is cached now, and re-measured only when the window size or the zoom changes.
- The problem report measures stalls too. The diagnostic file only had frame statistics; a 200-millisecond freeze was written nowhere — meaning the bug fixed in this release would not have shown up in a diagnostic even if it had been reported. The count, total and longest of tasks that hold the main thread for more than 50 ms are now recorded.
3.9.326
- Subtitles now appear at the same moment as the speech. The text used to reach the screen seconds before the sentence was spoken; measured, at its worst it was 7.4 seconds early. Subtitles appearing while nobody was speaking made up 19%; that is zero now. When the speaker stops, the text goes too.
- The Karaoke and Comic Pop styles are fixed. Both appeared word by word in the preview but came out as plain text in the video. In Karaoke the words are now coloured in as they are spoken, and in Comic Pop each word appears one at a time in its own colour.
- The Stop button really stops. Cancel during a download and the app carried on until the download had finished — on an 8-hour video that was the same as not cancelling at all. It now stops within half a second, and the half-finished file does not keep downloading in the background.
- « Remove the music / isolate the voice » works now. The installer was silently skipping a missing component and the feature threw an error. The installer now tests that the component really runs, and completes it itself if it is missing.
- The History tab became usable: the « Clear all history » button was moved above the list (down at the bottom it was hard to reach in a long list), and « Copy URL » was replaced by « Open the video » — one click takes you to the video's page.
- Two settings that did nothing were removed: HuggingFace Token and Fast download with Aria2. Both sat in the interface doing nothing at all; the first could even block a download while it was ticked.
3.9.324
- The app updates itself: when a fix is published, « An update is ready » appears on screen, and one click downloads and applies it. For most fixes you no longer need to reinstall the app.
- It tells you when a new release is out: if the app itself has changed, a « A new version is ready » card appears; it downloads and installs the setup file itself and reopens the app. No need to go looking for the website.
- Only the changed files are downloaded — you do not wait for a whole package for a small fix.
- Every update is verified by signature; if anything goes wrong, the version you have stays exactly as it is.
3.9.318
- Installation is a single piece: the installer brings everything needed — no manual steps left.
- No extra downloads after installation; progress is counted step by step.
- Technical component names were hidden from the interface and the installer.
Show older versions
Older
- The move to a desktop application was completed: an embedded sign-in window, a cookie bridge, the job engine.
- A distributable installer was produced — « next, next, install » became real.