2026-11-22 · mkok
Making a lyric video, the part that eats the time is never the look. It is the sync: eight words in a line, each lighting up on its own beat, is a full day of dragging markers — and one change to the arrangement means starting over.
mkok takes a different route: it does not listen to the song, it reads the score. Give it a lead-sheet MIDI exported on its own from the arrangement, and every note already carries an exact start and length. Line the lyrics up in order and the whole problem becomes counting.
The other approach is speech recognition plus forced alignment: let a machine hear what is being sung, then bind the lyrics to the sound. It sounds cleverer, but singing is a hard case — held notes, harmonies, reverb, a vocal that swells and fades will all make it bind in the wrong place, and once it is wrong, finding which line is wrong is its own afternoon.
MIDI has none of those problems. When you wrote the song, the melody was written as notes: at 3.2 seconds the note is “bed”, at 3.75 it is “fore”. The timing is in the score already; there is nothing to reverse-engineer from the sound.
The cost is explicit: you need a lead-sheet MIDI. What that buys is strict one-to-one correspondence — when it does not line up, it says so, rather than quietly fudging.
The lyric file is a plain txt. The rules are few:
The same lyric file lays out both a horizontal and a vertical version; you do not write it twice.
The algorithm fits in one sentence: sort every note on the chosen track by time, then let each lyric unit collect the notes it needs, in order.
A character highlights from the start of its first note to the end of its last — not a fixed duration, and not “until the next character”. So a held note naturally sweeps slowly and a fast passage sweeps quickly.
Take too many notes or too few and it reports an error on the spot, precisely: which line, which character, how many notes are needed, how many are left. That is deliberate. A misaligned subtitle is far harder to find than an error message.
When does a line appear? It fades in 2 seconds before its first note, soon enough to read, but no earlier than 0.5 seconds after the previous line ends; it disappears 0.5 seconds after the last note. A long interlude clears the line rather than leaving the previous one sitting greyed out.
The highlight progress is a fraction from 0 to 1: time elapsed ÷ the duration of that character. To draw it, the whole glyph is treated as an image and cut at “fraction × width”: lit colour on the left, unlit colour on the right.
The cut itself is feathered over about a pixel, using a very narrow grayscale gradient mask — so zoomed in it is not a hard line but a band of light with thickness.
The glow is the same idea stacked: offset the glyph, recolour it, blur it, and layer it a few times (3 by default). Flicker and breathing just move those layers' intensity over time. The particles are fully deterministic too: each character's particles come from a fixed random number derived from its position and the particle index, so exporting the same project twice sends the particles along exactly the same paths — no “this one looked better than the last one”.
The common failure in tools of this kind is “looked great in preview, fell apart on export” — because preview used one renderer and export used another.
mkok has a single renderer (Core Text + Core Graphics — no Metal, no Core Image), and preview, alpha export and burned-in export all call it. That is not an optimisation, it is a structural guarantee: if the preview is right, the export cannot be wrong.
One more detail: export takes a deep copy of the style at the moment it starts. Otherwise changing the font size during a ten-minute export makes the first frames and the last frames disagree. Being able to change settings mid-export is a trap, not a feature.
Exporting the subtitles alone as ProRes 4444 with alpha is the capability mkok is most asked for: drop them straight into an edit, no keying.
There is a trap here. The pixels the system renders are premultiplied — colour already multiplied by alpha. Hand that straight to the encoder and the colour of every semi-transparent edge gets darkened a second time, which shows up as a black fringe once it is composited over anything. So before each frame is written, vImage has to un-premultiply it. Using the system's optimised implementation costs 4.68 ms per frame; a hand-written loop costs 41.33 ms — nearly ten times more, which on a ten-minute piece is minutes of waiting.
And export is not done when the file is written: mkok reads the result back, takes the fourth byte (alpha) of every frame and builds a histogram, and passes only if at least one of three samples really contains transparent pixels. “We exported alpha” and “we exported alpha” are two different sentences.
.mov (ProRes 4444) is the master: visually lossless, audio passed through as far as possible. It is also enormous — about 253 KB per frame for subtitles alone, about 1.36 MB per frame with a background video, more than five times as much. A ten-minute master with a background is 18,000 frames and a very normal 20-something GB.
.mp4 (H.264) is the sharing copy: 12 Mbps, audio always re-encoded to AAC 192k. There is a limitation worth admitting up front — an mp3 track cannot go into an mp4; forced in, it plays as silence. So mp4 always re-encodes the audio, and only .mov can talk about passing it through. If you want genuinely untouched audio, export .mov.
A size estimate sits under the format menu, so you know how much disk the export will take before you start it.
No automatic lyric recognition. Speech recognition on singing is not accurate enough for a whole video, and one wrong spot means hunting through the whole thing. Requiring a MIDI is the honest trade: it raises the bar rather than offering a shortcut that has to be redone.
No uploads. The entire pipeline runs on this Mac. The only network communication is the Mac App Store purchase.
The sample project exports free. The built-in demo exports as it is, no purchase needed — try the formats, the sizes, the workflow; see the output before deciding whether to buy.
A Windows build ships as its own separate installer, released on its own page.
Mac App Store: id6819608702 · macOS 15.0 or later · Free download; export is a one-time purchase, yours forever, no subscription