Motion, Sound and Music

Ritu and Anand had five days and forty thousand rupees to make a five-second logo sting. It would run at the head of every branch video and on the idle screen of a hundred and forty ATMs, so it would be seen by the same tellers several hundred times a year. The contract included two revisions. A third would be at the studio’s cost.

The first cut was competent. The logo slid in from the left over twenty frames at twenty-five frames a second, the tagline faded up underneath, the colours matched the brand sheet and the typography had already been approved by the bank’s agency. The marketing head watched it twice and said it looked cheap. Asked what he meant, he could not say. Asked what would fix it, he could: motion blur on the logo, a light sweep across the letters, and a few drifting particles behind them.

Anand could do all three in an afternoon. He had the presets.

Ritu opened the graph editor instead, which she admits she did not usually do. The speed graph was a rectangle. The logo covered exactly twenty-four pixels on every one of the twenty frames, went from nothing to seven hundred and twenty pixels a second in no time at all, and stopped dead. The tagline started and stopped on the same frames as the logo, so the two elements moved as one rigid block.

That gave them a decision with three days left, and both answers cost something.

Doing what the client asked was four hours. It was visible, it was pointable-at in the next review, and it would have produced a floaty animation that was now slightly soft. Motion blur exists to stop fast motion strobing and to match synthetic motion to live footage; it does not create mass, and neither does a glow.

The other answer was to tell a paying client that the three things he had asked for would not fix his problem, and then to hand back something that looked, in the timeline, almost identical to the version he had rejected. The changelog would say “adjusted easing”. There would be nothing to point at. If it came back a third time, the studio would be working for nothing against a live date.

Ritu re-spaced the move: a four-frame anticipation in which the logo scaled down to ninety-six per cent, most of the travel in the first third, an overshoot to about a hundred and six per cent arriving fast, and a settle back to a hundred taking two and a half times as long as the overshoot. She offset the tagline three frames behind the logo and let it arrive after it. Nothing else changed: same colours, same type, same five seconds.

Then she hedged. She rendered three versions: the original with the requested effects, the re-spaced version with no effects at all, and the re-spaced version with the effects on top. She took all three to the review and did not say which was which.

There was one more decision buried in this, and it was about where the sting would live. On a branch video it is watched once. On an ATM idle screen it loops in front of a teller for six hours a day. Anticipation and overshoot are exactly the principles that make a single viewing memorable and the four-hundredth unbearable, so Ritu built the ATM cut with the same spacing and the overshoot reduced from six per cent to two, small enough to be felt rather than seen. Two files, one decision, and it added twenty minutes.

The app does one thing that matters. An invigilator taps a batch, a sheet rises from the row they tapped carrying that batch’s student list, and they mark absences before the test starts. The sheet enters over three hundred milliseconds on an ease-out, the rows arrive with a forty-millisecond stagger, and a scrim dims the list behind. On the team’s laptops, and on the one recent phone in the office, it is pleasant.

The centres reported that the app hung when you opened a batch. Neither developer could reproduce it. The tickets kept coming, and one centre head said his staff had gone back to paper for the morning slot.

They borrowed two of the phones the invigilators use, four-year-old Androids on the cheapest tariff, recorded the screen at sixty frames a second, and counted frames from the finger touching the row to the student list being readable. Seven hundred and eighty milliseconds. The animation they had budgeted for was three hundred.

The extra came from three places, and none was the animation being slow. The network request for the student list started when the animation finished rather than when the finger landed, which added two hundred milliseconds that were purely arithmetic. The row stagger was uncapped, so on a batch of twenty-four the last row began nine hundred and twenty milliseconds after the first, and although the top rows were readable early, the sheet was not interactive until the whole set had settled. And the sheet carried an animating box-shadow across its full width, which has no layout cost and repaints every frame, over a large area, on a weak GPU.

One developer proposed deleting every animation in the app. It was an afternoon’s work and it would measure fastest.

The other pulled the support log. Before the sheet existed, the commonest call had been “I marked the wrong batch”, and fixing one of those means a person editing rows in the database and, twice, driving to Kota. The sheet growing out of the row the teacher tapped is what tells them which batch they are in; a sheet that fades in at the centre of the screen is a new, unexplained object. Removing the motion would trade a complaint about speed for a complaint about data, and only one of those costs a day.

So the decision was between an afternoon with a known regression and a week of measurement and re-engineering, on a two-person team, with the exam season four weeks out.

They took the week. The request now starts on touch-down and the content arrives during the transition. The stagger is capped at six rows, with the rest appearing together, and every row is tappable while it is still animating. The shadow moved to a pseudo-element whose opacity is animated, which is a compositor property. They also added a reduced-motion path, because one centre head in his fifties had said the app made him feel “swimmy” and nobody had thought to write that down as a bug.

The reduced-motion work was smaller than either developer expected, because they had put their durations and curves in four custom properties rather than scattering them through the components. Redefining those four inside the media query, rather than overriding each screen, meant the behaviour reached parts of the app neither of them had opened in months, including two screens written by a contractor who had never heard of the setting.

Meera animated the upload loader in After Effects: three dots, a soft glow, a blurred trail behind the leading dot, a two-second loop. She exported a 1080p MP4, six hundred and forty kilobytes, and sent it to Sathish with the project files.

Sathish could not use it, and the reasons were not about quality.

The portal has a dark theme that buyers in Europe switch on and a light one the mill office uses. The loader has to sit on three different surface colours. An MP4 has a background, so it can sit on exactly one. The loader also has to stop on a “done” pose when an upload finishes, rather than loop until it is hidden; a video plays and takes no instruction beyond play and stop. Its colours are baked, so it cannot follow a theme. And it is a fixed size, so on the wide desktop layout it is soft.

That is four separate failures and none of them is a criticism of the animation. The animation was fine. The file was wrong.

They had two days before a customer demo, and three options.

Meera could rebuild it as a Lottie. The glow is an effect and effects do not export, so it would have to be reconstructed as shapes with gradients. The trailing blur is also an effect. The dot timing came from an expression, which either does not survive or is baked into hundreds of keyframes on export. Realistically two days, which was the whole margin, and she had another job booked for one of them.

Sathish could convert the MP4 to a GIF and ship it. Three hours. It would be around three megabytes on a portal whose users are frequently on mobile data, and because GIF has one bit of transparency the anti-aliased edges would carry a hard white fringe on the dark theme: the halo problem, which cannot be fixed inside the format. It would loop forever, so he would hide it and show a tick instead of resting on the done pose.

The third option came up on the third phone call, from Sathish. Three flat dots with staggered opacity is twelve lines of CSS. It inherits the theme’s colour because it uses currentColor, it weighs nothing, it scales, it stops whenever he wants, and it took an hour. It is not Meera’s animation.

That was the uncomfortable part, and it was not really about pride. Meera’s objection was that the agency had been paid for a designed loader and was about to ship a generic one, and that the next request would be for the developer to “just do it in code” every time. Sathish’s objection to the Lottie was that the file would arrive with four layers called Shape Layer 14 and no note about which frames meant what, as the last three had.

Both objections were about handover rather than about format, and both were fixable in an afternoon of writing things down. The format argument was the one that had eaten two days, and it turned out to have an answer neither of them had asked for out loud: the loader and the empty-state illustration were two different jobs that had been given the same file type because one person owned both.

Three chemistry lecturers, forty minutes of usable takes, and a plan to key each of them over animated diagrams for a revision series. The shoot had been cheap on purpose: a green cloth pinned to the wall of a seminar room, two phones, one recording, one as backup, and the room’s own tube lights.

The assistant keyed the first take in DaVinci Resolve’s free edition. On a paused frame it looked clean. The moment Dr Kulkarni turned her head it fell apart: stair-stepped edges through her hair, a green rim along her shoulder that survived every setting, and a matte that crawled between frames even where nothing moved.

Three causes, all of them decided before the camera rolled. The phones recorded 4:2:0, so the colour information a keyer builds its matte from exists at half resolution horizontally and vertically: the colour edge is two pixels wide where the luminance edge is one, and no amount of edge refinement recovers what was never sampled. The cloth was lit by the room, so it was about two stops brighter on the left than the right and no single tolerance covered both. And the lecturers stood roughly sixty centimetres from the cloth, so green bounce landed on the side of every face and shoulder.

The break was half over when they had to decide.

Reshooting meant borrowing two mirrorless bodies from the photography department that could record 4:2:2 to an external recorder, hiring two soft lights for four days at about nine thousand rupees, and getting three lecturers back: two of whom had gone home to Nagpur and Belgaum, and one of whom had made it clear that he had done this as a favour and would not be doing it twice. The cost was a week the cell did not have, money it did not have, and goodwill it could not replace.

Working with what existed meant a free segmentation model to generate a garbage matte and a core matte in minutes per shot, hand-roto confined to the edge band between them, despill, and a light wrap built from the diagram behind. Estimated at three weeks of the assistant’s time for forty minutes of footage, and the hair would still boil.

On the second day the assistant proposed a third thing: stop compositing. Put the lecturer in a corner of the frame at a fixed size, give the diagram the rest of the screen, and key nothing at all. The green cloth then stops being a matte source and becomes a background to be graded to the series colour, which it is perfectly good at.

The head of the cell resisted it at first, and his reason was not vanity. The series had been pitched to the principal as “the lecturer inside the diagram”, and a corner box looks like every recorded Zoom class the students had already sat through for two years. Against that, the assistant put a number on the alternative: at three weeks for forty minutes, the series would not exist before the exams it was made for.

They also checked the one thing that might have rescued the original plan. The backup phone had recorded the same takes from a slightly different angle, and if either device had been set to a higher bitrate the chroma might have held up better. It had not; both files were the same 4:2:0 at the same rate. There was nothing to go back to, which is the ordinary state of affairs when the shoot is discovered to be the problem.

Two hundred videos, eight to twelve minutes each, graded in Resolve and exported as H.264. The upload took a fortnight on a ten-megabit line. The portal’s reviewer rejected the whole batch in one line: washed out, low contrast, the blacks are grey.

On every machine in the Bhopal office, and in VLC, the files looked correct.

The cause took an hour to find and cost nothing to state. The export had written no colour tags at all: no colr atom, no range flag, nothing in the bitstream’s video usability information. The files were limited range, with black at sixteen and white at two hundred and thirty-five. The reviewer’s player, faced with an untagged file, assumed full range and treated sixteen as a real sixteen, so every black sat above zero. Milky means limited data read as full. Nothing in the files was damaged; the numbers were right and the description of them was missing.

Three ways forward, none free.

The first was to re-tag. ffmpeg -c copy with the four colour flags writes the description into the container with no decode and no re-encode, about two minutes a file, and no generational loss at all. The processing was trivial. The re-upload was not: another fortnight on the same line, three weeks before term started.

The second was what the producer wanted. Take the reviewer’s screenshot as the truth, add contrast, crush the blacks until it matched his player, re-export everything. Two days of machine time and the same fortnight of uploading. It would pass the review. It would also make every file wrong in every correctly-behaving player, including the portal’s own Android app, which honours the tags: the same footage would then arrive crushed in the shadows and clipped in the highlights for the students it was made for.

The third was to argue: send a MediaInfo readout, ask the reviewer to open the file in a second player. Free, and it puts a four-person supplier in the position of telling a state portal that its review is wrong.

There was a fourth thing they considered and rejected, which is worth recording because it is the one most teams reach for. They could have re-exported with the tags written, rather than copying the existing streams. It would have been the same upload and roughly the same calendar time, and it felt tidier: one clean pass, one set of settings, everything consistent. It would also have put two hundred already-compressed lessons through a second generation of lossy encoding, spending real quality on a problem that was not in the pixels at all. Every encode quantises, and the second encoder would have been working to reproduce the first one’s block edges as if they were detail.

The argument inside the office was not about the physics, which nobody disputed. It was about whose fault the rejection looked like. Sending a corrected file with an explanation reads as a supplier fixing its own error; sending a MediaInfo readout reads as a supplier correcting a client. They wrote the note four times before sending it, and the version that went out did neither: it showed the tags missing and present, said what each player would do with each, and asked him to try a second player before they re-uploaded the remaining hundred and ninety-nine.

They did the first and the third together, in that order and on one file.

Manpreet and Iqbal record fifteen-minute revision talks. They go out on memory cards to students without data and on WhatsApp to those who have it. The first twenty were recorded on a phone laid on the staff-room table, about a metre and a half away. Students said they sounded far away. One boy said he could not follow them at all on his father’s phone speaker in the shop.

Then a district innovation grant arrived: twelve thousand rupees, to be spent on equipment, itemised in the register. A vendor quoted a USB condenser microphone and a desk stand. The headmaster liked the idea of a microphone. It is a thing, it can be shown to a visiting officer, and it is unambiguously equipment.

Manpreet measured before buying anything, which took about twenty minutes. She recorded ten seconds of room tone in the staff room and ten seconds of speech, and compared them with ffmpeg’s astats: the voice sat twenty-eight decibels above the room. She moved the phone to twenty centimetres and measured again: forty. Then she recorded the same paragraph three ways: staff room at a metre and a half, staff room at twenty centimetres, and the store room at twenty centimetres with two spare curtains hung on the facing wall and a folded blanket under the phone. She played all three on the cheapest phone in the building.

The store-room version was not slightly better. It was a different recording.

That produced an awkward decision. The two changes that had done the work (getting close, and putting soft mass on the surfaces facing the microphone) cost nothing and could not be entered in a register. The purchase that could be entered would not change the direct-to-reverberant ratio at all, because that ratio is set by distance and by the room, and the microphone’s price is not on the list.

Manpreet asked the office whether curtains and mineral-wool panels could be booked as acoustic material. The office said no: soft furnishing is not equipment.

So they split it. Three and a half thousand went on a wired lavalier and a small handheld recorder, so that a teacher could record in the store room without carrying a laptop. Twelve hundred went on a second memory card and a backup card reader. The remaining seven thousand bought the condenser the headmaster wanted, and it lives in the store room, used at twenty centimetres, behind curtains the science department donated because they were being replaced anyway.

Manpreet was clear with herself that this was not a pure outcome. Some of the grant went on the least effective item in the room, and she agreed to that knowingly, because a grant returned unspent is a grant the district does not offer again and because the headmaster’s support was worth more over a year than two thousand rupees of equipment. What she refused to do was let the purchase stand in for the change: the microphone was bought, and the recording position moved anyway.

The second thing she insisted on was cheaper still. Every talk now begins with thirty seconds of the room, recorded at the same gain, in the same position, before anyone speaks. It sounds like nothing. It is what lets a sentence be cut out of the middle of a talk without the background stopping for a frame at the join, and it is impossible to go back for once the chairs have moved.

The film announces a free screening camp: where it is, which days, who should come. It goes out on WhatsApp and is played by ASHA workers on their own phones, frequently held up in front of two or three people in a courtyard with a fan running.

Sneha mixed it the way the module says to. Dialogue anchored and consistent, the music bed sitting about seventeen decibels under the speech while anyone is talking, integrated loudness at minus sixteen LUFS, true peak at minus one point five.

The communications officer listened on the studio monitors at a comfortable level and said the music felt thin and unimportant. He asked for it up six decibels. On those monitors, in that room, he was right: it did sound thin.

Sneha checked the same file on a six-thousand-rupee phone at arm’s length, at low volume, with the studio’s own fan switched on. At the officer’s level the Marathi voiceover was hard to follow, and the reason was specific. The bed had a bright string line living between one and a half and three kilohertz, which is squarely where speech intelligibility lives. The phone produced almost nothing below about six hundred hertz, so everything that had given the bed its body and weight on the monitors was simply absent: what survived on that device was its midrange, and its midrange was the voice’s.

The audience made it worse rather than better. A large share of the people who would watch this were over sixty, and the ability to separate speech from a competing sound falls with age and with hearing loss. Complaints about background music drowning speech come disproportionately from exactly those listeners.

Three ways to handle it. Do as asked, sign off today, ship on Friday, and lose the dates for the people most likely to need the camp. Refuse and explain, which is a supplier telling a government client that his ears are wrong with a launch four days out. Or change the arrangement rather than the level.

She took the third. She cut three decibels with a wide bell at two kilohertz in the music bus, which takes away the part that was in the way and almost none of what the bed is actually made of. She raised the bed four of the six decibels he had asked for. She put a two-second hole in the music under the single line that carries the camp’s dates, edited into the arrangement rather than ducked by a compressor. And she took the review to the phone.

She had tried a ducker first and taken it off again. A sidechain can only react after the fact, so it was pulling the bed down on every breath and every lip smack and letting it surge back between sentences: the pumping that makes an automated intro recognisable as one. Riding the level by hand meant she could start the move two hundred milliseconds before a line began and release it slowly, which is what a human can do and a detector cannot. On a ninety-second film there were eleven moves to draw, and it took less time than tuning the ducker had.

The deeper problem was never the level. Choosing a bed with a bright string line sitting in the intelligibility band had made every later decision harder, and a track with sparse midrange and its energy below five hundred hertz or above six kilohertz would have needed almost none of this work. Sneha now listens to candidate tracks on the phone before the edit is locked, which costs two minutes and has changed which tracks get chosen.

Twenty of the sixty lessons had been narrated by Ms Beena, a teacher on deputation, in four hours and twenty minutes of clean, close-miked recordings. Then her deputation ended and she was posted three hundred kilometres away.

Three ways to finish. A new narrator, which changes the voice in the middle of a series that children would work through in order. Bringing Beena back for four Saturdays, at an honorarium of eight hundred rupees a day plus travel she would have to arrange herself. Or cloning her voice from the material already recorded.

The technical assistant built a test clone and played it to three colleagues who knew her. At three seconds all three said it was her. At thirty seconds all three said something was wrong and none could say what: the timbre was right and the timing and stress were somebody else’s, the generic behaviour of the synthesiser’s training distribution. On one line with a contrast in it, the stress landed on the wrong word, in a completely convincing voice.

The director insisted the decision be taken as three separate questions rather than one.

The first was consent. Beena said yes on the phone in under a minute. The director asked for it in writing and scoped: for which project, for which uses, for how long, and whether the trained model itself would be kept, where, and who could run it. Beena agreed to the remaining forty lessons of this series, for three years, with the model held on the institute’s own machine rather than a service and deleted at the end. She declined a clause offering the voice for future departmental material, the clause she said afterwards she would never have noticed in a one-line permission.

The second was quality. The assistant built a pronunciation lexicon of about sixty recurring terms (element names, three place names, the word for precipitate) which removed that category of error permanently. He then marked stress by hand in any sentence containing a contrast, because the written text does not encode stress and stress is the meaning. And every lesson would be listened to end to end against the script before publication, about forty minutes a lesson: twenty-seven hours of listening across the series, budgeted in advance rather than discovered.

The third was disclosure. India’s amended IT rules for synthetically generated information prescribe a label a person can perceive, and for audio one in the opening portion. They recorded a single line at the head of each lesson (in Beena’s own real voice, cut from the existing material) saying that the narration that follows is a synthesised version of her voice, used with her permission.

That line was written before a single lesson was generated, which the assistant now says was the only reason it exists. Written first it is one sentence. Written after forty lessons have been rendered and timed against slides, it is a change to forty locked timelines, and it does not get made.

The costs were not evenly distributed, and the institute counted them honestly afterwards. Bringing Beena back for four Saturdays would have been about six thousand rupees in honorarium and perhaps three weeks of calendar time waiting on her availability, with no consent question at all and no listening pass beyond the ordinary one. The clone cost five weeks of an assistant, twenty-seven hours of listening, a written agreement somebody had to draft, and a deletion to remember eighteen months later. What it bought was a series that sounds like one series, finished before the exam term, in the voice the first twenty lessons had already established.