Video Cleanup

The camera file is converted once into a working master: 1920×1080, square pixels, a constant 30 frames a second (the studio cameras record 59.94 fps with uncompressed audio, which nothing downstream wants), high-quality video and 192 kbit/s audio. It is written under a temporary name and moved into place only when complete. On every later run its length is compared with the original, and if they differ by 2 seconds or more it is treated as truncated and made again. A half-written file that looks finished has caused real faults before.

A 16 kHz mono copy of the sound goes through speech recognition in chunks of about 24 seconds. Each chunk boundary is placed on a real pause where one exists: a silence of at least 0.28 s below −28 dB, chosen between 12 and 32 seconds after the previous boundary. Where no such pause exists the boundary falls at the 24-second mark. Word pieces are joined into words, each word gets a start and an end time, and no word is allowed to overlap the next or last less than 60 ms. Every decision after this point is read from the transcript and its timestamps, not by ear.

An automatic pass reads the transcript for speech that is not the lecture. It splits the speech into islands, meaning runs of words bounded by at least 1.10 s of silence, and scores any island of 14 words or fewer. Recording talk such as 'are we rolling', 'can you hear me' or 'let me take that again' scores 3. Words about the recording addressed to someone in the room score 2. An island standing outside the first or last line across a pause of 2 s or more scores 2, and one simply inside the first or last 25 s scores 1. A short question of 6 words or fewer on its own scores 1. An island is cut at a score of 3. Slide headings read from the script ('slide two', 'next slide', 'slide three of nine') are matched wherever they occur, even mid-sentence, along with the short verbless title lines around them.

Retakes, false starts and stutters are correct English in the middle of correct English, and only the meaning shows that they are a second attempt. The engine does not detect them automatically. They come from a hand-written cut list, prepared by a person reading the transcript or by the script-repair pass. Each entry names the first and last word index, a category and the quoted words. The build refuses a cut list that points at words outside the transcript.

Every word-level cut is widened by 50 ms either side so a plosive or a breath next to it is not sliced. The head is trimmed to 0.45 s before the first surviving word and the tail to 1.00 s after the last. Both are measured from the words that survive, so a mic check before the first sentence takes its surrounding silence with it. Any pause between surviving words longer than 1.20 s is cut from its middle down to 0.50 s. Overlapping removals are merged, and the stretches left between them become the kept segments: the master timeline.

Every surviving word is moved onto the new, shorter timeline by subtracting everything removed before it. A word squeezed to nothing gets 60 ms, and overlaps are clamped to the next word's start. This renumbered transcript is what captions, the animated deck, the cutaway planner and the bookends all read. After spelling repairs from the lecturer's own script, it is grouped into caption phrases: up to 9 words for the band, 6 for the column, broken at punctuation, at any gap over 0.42 s, or at 3.6 s of running phrase.

The voice track is cut in one pass directly from the plan, with a 20 ms fade in and out at every join so no cut clicks. It comes out as mono, 48 kHz, 16-bit, and it is the only input Studio Sound ever sees.

The picture is cut one kept segment at a time and the segments are joined afterwards, because a single edit across 25 minutes of 1080p is slow and will be killed for memory. Each segment is found with a two-stage seek: a coarse jump to 4 seconds before the cut, then an exact seek from there. It is ended by frame count rather than duration, so rounding cannot build up across segments. Each segment is then measured. One that misses its length by more than 1.5 frames stops the build. Each segment is stamped with the start and end it was cut from and reused on a later run only if the plan still asks for exactly that. Segments from an older, longer plan are moved aside, not joined.

On the cut picture, the green backing's real colour is measured across 9 frames and the key's reach is chosen by trying it. The presenter survey then keys 48 frames spread across the whole lecture with that same key, ignores the outer twentieth of the frame on each side, and keeps the person rather than the scenery. It records a wide box that holds every gesture, a tight box for the small panel, and the median centre of the head. These measurements are written once and read by every presenter layout.

The last step of ingest is the job doctor. It checks that the cut sound matches the plan's length, that the cut picture matches the cut sound to within 2 frames, and that every segment is stamped with the current plan. It also checks that the key and the presenter survey were both measured on the cut footage they will be used on. A failure stops the job before any deck, plan or render is made against a timeline that is wrong.