Presenter + Captions

The transcript's timing is kept exactly, but the recogniser's spelling of names and technical terms is corrected from the lecturer's own script. Words are then grouped into caption phrases for the column. A phrase breaks after 6 words, at a full stop, question mark, exclamation mark, comma, semicolon or colon, where the next word follows a pause of more than 0.42 s, or once the phrase has run 3.6 s. A single word is never left as a phrase unless a pause or the end of the lecture follows it. The bottom caption band uses the same rules with a 9-word limit. The column allows 6 because it is narrower.

Nine frames are sampled between 5% and 95% of the recording. A pixel counts as backing when its green exceeds red by 40 and blue by 60 and is above 90. The key colour is the median of those pixels. The key's reach is then tested, not guessed. For each value on a ladder from 0.06 to 0.24, the real matte is built with the same keyer the render uses and scored on two questions: how much certain backing survives, and how much of the certain presenter is kept. The winner is the smallest reach that leaves no more than 0.05% of the backing while keeping the presenter within 0.2% of the best score. One further rung of headroom is taken only if it costs none of the presenter. Edge softness is 0.06 and green spill is neutralised at a 0.5 mix.

Forty-eight frames spread across the whole recording are keyed with that same key. The outer 5% of the frame on each side is ignored, because the strip of studio wall past the edge of the cloth keys as foreground. Shapes that overlap the main figure's columns count as the presenter. Shapes off to the side count as scenery. The box's sides come from the 4th and 96th percentiles of every sampled frame, its top from the 5th and its bottom from the 98th. It is then padded by 1.2% of the frame width and given 6% of its height as headroom, because hair keys out thinly. The engine also records the horizontal centre of the head, which is not the same as the centre of the gestures.

Before any layout is chosen, the planner runs the sizing test. It looks for the widest presenter panel, from 1200 px down in 10 px steps, that leaves a gutter (70 px for panels over 1000 px, otherwise 90 px) and a caption column of at least 670 px inside x = 1860, and that puts the top of the head at or above y = 270. The formula is head_y = 1080 − union height × panel width ÷ union width. The panel width, the column and head_y are written into every Presenter + Captions run in the plan. If no width passes, the layout is removed from the menu and the numbers that ruled it out are reported. A wide, gesturing presenter changes how the layout is sized. That alone never removes it.

The composition planner gives the layout the lecture's opening, up to the phrase boundary nearest 15 s. After that, Presenter + Captions competes with the other layouts on a score of 0.55 × the delivery score plus 1.4 × the definition score. It is also used to cover a stretch where the animated slide shows nothing but its heading for 4.5 s or more, and to break an animated run that would otherwise pass 60 s. In both of those jobs it alternates with Presenter Centred. Any stretch it wins is cut back to 15 s, and the rest goes to the next layout that stretch preferred.

The renderer draws one full-length plate: the stage gradient, the caption column, the chapter heading above it, and the timer and progress bar along the bottom. Where the presenter will stand, the plate is left empty. Because it is drawn for the whole runtime, any stretch the plan chooses can be cut from it.

For each Presenter + Captions stretch, the assembler cuts the same time range from the plate and from the cleaned footage. The footage is keyed and despilled, the shadow side is lifted, and the result is cropped to the measured presenter box. It is scaled so the figure is 1004 px tall and placed flush with the right and bottom edges of the frame, which puts the top of the box at y = 76. The box's left side is keyed backdrop, so it can sit near the column without hiding anything.

A cutaway scheduled inside a Presenter + Captions stretch replaces the whole frame for its length and then hands back to the presenter. The caption band is burned over the cutaway, so the words carry on across the picture. A drawn cutaway placed over this layout is given the Floating & Pulsing Glow treatment, unless that would repeat the previous drawn cutaway's treatment.

Every piece is checked against the frame count its duration requires, within 2 frames, before anything is joined. A piece that encodes at the wrong frame rate still has the right duration and would pull everything after it out of sync with the audio. Each piece also records fingerprints of the footage, key, crop and look it was made from. It is rebuilt, not reused, if any of them has changed. The studio voice track is added once, over the whole film, at the end.