Release history

Changed

• Internal developer diagnostics were updated; nothing changes in how the panel works for you.

Changed

• "Plans" opens the same page, tagged as coming from the plugin.
• Five tips now say what the panel does: Count Who Speaks no longer promises a transcript, the lead before a line is 0.25 s, Who Speaks is about cameras and tracks, and Breaths and "Um" and Pauses Only in Dialogue describe what actually happens.

Changed

• Updated the API.
• Short format picks the strongest moments of the whole recording the way editors do: it weighs every option found, not just the ones that agreed, and more often builds a clip from different parts of the recording.

Changed

• Editing progress now shows the time for each step: "~12 min" on the running one and on the ones waiting. If a step takes longer than promised, it says "soon" and the fill keeps creeping forward. The Model Analysis tile has three dots: reading → outline → edit.
• The bottom bar and the line next to Stop are gone.
• The first estimate now takes your computer into account: without an NVIDIA card, sound analysis is estimated from the number of processor cores, not from a graphics card, and is corrected by the real speed after half a minute.
• If the model analysis fails, the panel no longer retries it again and again on its own; it says: "The analysis did not go through. Try again later — the work done so far is kept."

Changed

• Monologue now cuts with a new model: retakes across the whole recording, slips and junk. The model looks at a recording of up to 4 hours as a whole, and while it thinks, the line below shows what it is working on: "model is thinking: outline · 0:00–43:05", then "edit".
• After editing, next to "All Done" you see what was removed and why.
• Monologue on two cameras: the camera switches on splices and hides the jump, and inside a long piece only in a pause between words, never mid-word. No more shots shorter than a second and a half and no slivers left on the top track.
• The same number in "no more often" and "no less often" (for example, 6 and 6) is a rough rhythm, not a switch exactly by the clock.
• On computers without an NVIDIA card (including AMD or Intel graphics), speech recognition now uses the whole processor instead of half of it and runs 1.7–2.5 times faster: engine 1.1.0 works on several pieces of the recording at once. The text is the same. Nothing changes on NVIDIA cards.
• The panel remembers the cameras picked for mixing and the "no more often / no less often" numbers between launches.
• Under the progress on every tab there is now one overall bar and a line like "about 12 min left", so you can see how long you can step away. The estimate is refined as work goes and remembers how fast your computer is.
• A stage's bar no longer drops to zero on every sub-step and no longer stands still while the model analyses the first part.
• There is no more Standard / Strong model choice: paid hours always use the strong scheme, which catches re-shot takes and cleans up small slips. An hour of dialogue uses an hour and a half.
• The free part is now the first 30 minutes on the same strong scheme as paid hours, once per computer, instead of an hour a month. The button in the key window, the line next to the button and the hours panel say so: "Trial: 30 min".
• If this month's trials have run out, the panel says the pass will use a lighter model.
• Dialogue and Short Videos on paid hours now use the same strong scheme as Monologue. Dialogue gives the model the whole conversation instead of chunks: guests' lines no longer go missing at the seams, and while the model thinks, the line below shows what it is working on, just like Monologue.

Changed

• In Report a Problem, email and text are now optional; Show What Will Be Sent is gone.

Changed

• Dialogue: the recording is transcribed once — when you press Edit. Count Who Speaks only finds who speaks where, Next opens Pauses right away, and the pause and camera settings can be set without waiting. How much gets removed and how many shots there are is shown after the edit.
• Pauses Only mode — in both Monologue and Dialogue — no longer spends hours: cutting without the model is free. A key is still required; only model analysis spends hours.
• Dialogue now cuts like an editor: video only where the shot changes, at one point on all cameras; audio — each microphone only where it switches on or off. No idle cuts, and the edit in Premiere runs several times faster.
• Dialogue: the room microphone is switched off entirely when the speakers have their own microphones. It stays on only if there is one speaker microphone or none — then the other person can be heard only through the room microphone.

Changed

• Shorts: the strongest moments with a clear, catchy opening on any material; a video that starts mid-sentence is redone by the server. The empty prompt window shows short examples of how to steer the choice your way.
• Shorts from a finished episode: by default one video per 80 seconds of speech, up to 30 videos (was one per 2 minutes, up to 20).
• Everything Guillotine creates in the project is named in English whatever the panel language: the copy is "<name> Guillotine", shorts are "<name> Guillotine Shorts".
• In Who Speaks, once the speakers are counted, there is one button — Next: it transcribes the recording and opens the next step; if the transcript is already there, it opens it right away. The separate Transcribe Now button is gone.
• Under the pause settings you can see how much they remove and that the model does the main cut while editing; under the camera settings — how many shots the episode has and that they do not change the length.
• Shorts are now built from short meaningful pieces of 3–20 seconds — a sentence or a complete part of a thought — instead of long phrases of up to several minutes. A clip can combine thoughts from different minutes of the recording, and this also works on a finished episode with no pauses.

Changed

• A key entered on a new computer now moves there on its own — no need to unlink it on the old one first. A key can move up to 3 times in 30 days; after that, through support. The old computer says "The key was moved to another computer" at its next check.
• The panel header now shows the author line: "Guillotine by A.Sudorgin v<version>".
• The engine installer speaks English on any Windows except a Russian one; before, it was in Russian on German, French and other systems.
• Labels, tooltips and messages rewritten in plain language. Modes have the same names on every tab — Pauses Only and Pauses and Retakes; the tab is Short Videos, and the filler words box is now Breaths and "Um", after what it really removes. Every button and setting has a tooltip.
• Error messages say what happened and what to press, without codes, and point to Report a Problem. Units match Premiere: s, ms, dB.
• New panel look. Each format has its own color: Monologue amber, Dialogue light blue, Short Videos purple; switching formats recolors the panel smoothly. A new header shows the hours left with a live dot and "?", plus a new format switch with icons.
• Steps are cards: the open step is highlighted, completed ones get a filled circle, and the dotted line reaches each step as soon as you open it.
• Go back and the line returns to the open step while later completed steps keep their filled circles — the line shows where you are, the circles where you have been. Steps arrange themselves to the panel width: one column when narrow, two or three when wider.
• The Refresh button is gone because it is no longer needed: the panel sees which sequence is open and notices added and removed clips and tracks. With no sequence open, the panel calmly asks you to open one and picks it up by itself. Switch to another recording and the previous All Done clears.
• The same new look inside steps: track choice and "what to remove" are rows with a circle, checkboxes and fields are in the format color, and Edit and Build are large full-width buttons. In the short format the video length is one "30 — 120" row next to Number of Videos.
• Warnings (such as "the model did not analyze the recording") sit on a red strip so you cannot miss them.
• Edit progress is shown as tiles: each stage has its own tile that fills as work goes, with percentages next to it and "stage 2 of 3" on top. What is happening right now ("recognizing words", "3 of 8 windows analyzed") is a line next to Stop. On a narrow panel the tiles stack in a column.
• When an edit or a short-video build finishes successfully, All Done stays on screen: all tiles are filled, with "before → after" below them (a list of videos for short videos). It stays until the next run, a format change or a settings change.
• The panel opens with a key, like our other panels: a key window on first launch, then "?" → Key. The key field in the header is gone.
• Every mode on every tab spends hours the same way, by recording length: model analysis, silence-only cutting and Pauses Only in Dialogue alike. Short videos cost double.
• Work that failed (the model did not answer, the connection dropped, you pressed Stop) costs no hours: what was charged goes back to your balance. Retrying a window right after a failure is free.
• A trial run on the first 5 minutes in modes without the model spends 5 minutes, not the whole recording.
• The old pass is replaced with a key: the panel asks once for the key from your email. The previous panel version with the old pass works until November 1, 2026.
• Short videos match the size of your recording: 4K gives vertical 2160×3840, 1080p gives 1080×1920 as before. Short videos take the sequence color settings (working color space, linear color, tone mapping) from the source; if something does not apply, the panel says so.