Every release has a section with nothing to show. An intro with no vocal, a beat you made before the vocal exists, a drop, a producer tag, an instrumental B-side. Posting a static cover over it is what most people do, and the feed treats a still image with audio exactly the way it looks: as something to scroll past.

A visualizer is the format for that section. It is not a lesser lyric video. It is the right answer to a different question.

When a visualizer is the right format

Three situations, and they cover nearly every case.

The track has no words in the section you are posting. Instrumentals, beats, the sixteen bars before the vocal comes in. A lyric video of silence is a blank frame.

The words are not the point. A vocal chop, an ad-lib run, a heavily processed hook where the texture is the hook. Putting readable text over it draws attention to the one thing you did not want examined.

You have no footage and no time to find any. A lyric video over the wrong clip is worse than a visualizer over the right colour. The footage question can be skipped entirely here, because the motion comes from the sound.

If none of those apply and the section has a line people will sing back, make the lyric video instead. The two formats often end up posted for the same track, cut from different sections.

What makes one read on a phone

Four things separate a visualizer that holds attention from one that reads as a screensaver.

The motion is driven by the audio, not looped. A loop of bars bouncing to nothing is recognisable within two seconds. A viewer cannot say what is wrong, only that the picture and the sound are not the same thing. Every one of Revori's five styles reads the audio frame by frame and draws that, so a kick moves the picture when the kick happens and a breakdown lets it settle.

The backdrop means something. The visualizer is a layer, and what it sits on is most of the frame. The cover art is the usual choice and a good one: the listener sees the image they will later find in the app. A photo of you, a single flat colour from the artwork, a slow gradient. What does not work is a stock image chosen because it was there; the motion draws the eye and then the eye lands on nothing.

One style, chosen for the energy. A dense, bright style over a slow track is exhausting; a drifting particle field under a fast one looks asleep. Pick the style for the energy of the section, then leave it. Switching styles inside one post reads as indecision.

Room to breathe at phone size. A visualizer is watched in peripheral vision most of the time. Motion that fills the frame edge to edge on a desktop monitor becomes noise on a phone, and the outer band is under the platform's own overlays anyway. Preview at phone size before exporting, the same rule as for any lyric video.

The five styles

Each one draws the audio differently, which is the whole basis for choosing.

Circles. A ring of bars around a centre point, each bar driven by its own slice of the spectrum, the whole ring breathing with the bass. The centre is reserved for an image, usually the cover. Suits anything with a clear pulse, and it is the default because it frames artwork.

Orbit. A single orb that breathes with the low end and sends a ring outward on every beat of the tempo grid. It reads the kick more than anything else, so it suits tracks whose pulse is the point: minimal, techno, a beat where the drum is the hook.

Nebula Drift. A few hundred particles on a slow orbit, each tied to a band of the spectrum; the bass pulls the whole field inward and the highs make parts of it sparkle. Reads as texture rather than as a meter, so it suits tracks where you do not want a literal readout of the beat: downtempo, vocal-led passages, anything atmospheric.

Bars. The classic equaliser, weighted toward the low and mid bands so it moves musically rather than flickering. The most literal readout, and the right choice for high-energy sections where the listener wants to see the drop coming. Also the most legible at small sizes.

Waves. A mirrored wave across the centre of the frame, drawn from the mid and high bands rather than the kick, with a stroke that thickens on energy. It moves with whatever is out front, a vocal chop, a guitar line, a lead synth, and it is the one that looks most like the sound drawn.

If you cannot decide, Circles with the cover in the centre is the safe answer for a release and Bars is the safe answer for a drop.

What it costs and what comes out

A visualizer render is 40 credits, less than a lyric video render, because there is no lyric layer to draw. It needs the song upload but not the transcription: the visualizer flow skips lyrics entirely, so there is no transcript to correct and nothing to sync. The same upload is reused if you later make a lyric video or a Canvas from the track.

Output is 1080p MP4 in 9:16, 1:1 or 16:9, each a separate render because the motion is composed for the frame. Previewing and changing the style or backdrop is free and unlimited, so the right workflow is to try three styles against the section and export the one that survives the phone check.

"The lyrics are out of sync" is one sentence that describes four different problems. They look similar in a preview and they have nothing to do with each other underneath, which is why the first fix people reach for, dragging everything a bit earlier, usually makes one of them better and another worse.

Sort out which one you have before touching anything.

Which of the four is it

Play the section and watch for the shape of the error.

One word is wrong, the rest are fine. Alignment miss. Skip to the second section.

A whole phrase or a whole section drifts, then it recovers. Usually a transcript problem upstream of alignment, sometimes a passage the aligner could not hear. First and second sections.

Everything is off by the same amount, from the first line to the last. Global offset. Third section.

Everything is technically on time and it still reads late. The perceptual one. Fourth section, and it is already handled for you.

The transcript is wrong

A word that was never sung has no correct time. If the transcript says "fire" where the singer sang "higher", the aligner will find the best place for "fire" and it will be near the right spot, but the highlight will land on the wrong syllable count and the next few words shuffle to make room. It looks like a timing error. It is a text error.

This is why the transcription accuracy post says to read the transcript against the song before doing anything else. On real uploads the text is the bottleneck, not the alignment.

The fix is in the text, then a realignment. In Lyric Studio, correct the words, or paste your own lyric sheet if the transcript is wrong in more than a few places. Then run Realign: the text stays as you wrote it, the timings are replaced with a fresh alignment against the audio, and undo restores the old timings if it made something worse. Realigning after a batch of edits is cheaper than nudging twenty words by hand, and more accurate, because the aligner sees the corrected text as a whole.

One word or one phrase is off

The transcript is right and one word still lands wrong. This is the alignment's own error, and it has a shape.

Alignment pins each word to its acoustic onset by matching the text against the separated vocal. It is good at consonant attacks in a clear passage and worse at three things: a held vowel, where the "onset" is a slow swell with no edge to find; a quiet or sparse passage, where the vocal barely rises above the stem's noise floor; and stacked harmonies or heavy processing, where several voices start the word at slightly different times.

Since September 2026 the onsets come from a CTC aligner with a second opinion. Other timing estimators vote on each word, and where two of them agree with each other and disagree with the aligner by more than 200 ms, the word takes the vote. Where the estimators disagree with each other, the word is marked with a lower timing confidence and Lyric Studio shows a Timing chip on it. That chip is the list of words to listen to first.

On the JamendoLyrics benchmark, given the correct text, the median word is 47 ms from its true onset and 95 percent of words are within 300 ms. The remaining 5 percent are the ones you will notice, and on a three minute song that is a couple of dozen words. Nobody has an automatic system that removes that tail on sung vocals, so the design target is a tight bulk and a cheap manual fix.

Three manual fixes, from cheapest up:

Nudge. Open the word's popover. The arrows move it 50 ms earlier or later per click, 200 ms with Shift. Two clicks fixes most misses.

Tap. Play the song, and press T at the moment you hear the word. Its onset moves to that moment minus 120 ms, which is roughly how long a person takes to press a key after hearing something, and the word is marked as verified so the Timing chip and the issues list stop flagging it. Tapping works because your ear is exactly the instrument the aligner lacks on a held note.

Drag. In the Studio timeline, drag the word's pill along the waveform. Useful when the miss is large and you can see the vocal energy the word should sit on.

None of these touch the beat grid, and none should. Onset snapping, moving each word to the nearest detected vocal onset, was in the pipeline for six months on the theory that it was the accuracy lever. Measured against human labelled onsets it was making timing worse, and deleting it improved the median. The lesson generalises: a correction that cannot verify its target moves words away from the truth about as often as toward it.

The whole track is early or late

Every word is off by about the same amount, first line to last. Two common causes.

The audio changed. If the analysis ran on one file and the project now plays a different master, a version with a different amount of leading silence, or a re-encode with a different amount of padding at the start, every timestamp is shifted by that difference. The words are right relative to the audio they were aligned to and wrong relative to the audio you are hearing.

Or it is taste. Some people want the highlight noticeably ahead of the voice, in the way karaoke systems often run, and some want it dead on. Neither is wrong.

Either way the fix is one control. In the Style panel, the Sync card has a Lyric offset slider from 2,000 ms early to 2,000 ms late in 50 ms steps. It shifts every word together and applies to the preview and the export identically, because both are rendered by the same composition. Set it, scrub three or four places in the song to confirm the shift is constant, and stop there.

If the drift is not constant, if the first verse is fine and the last chorus is late, the offset slider is the wrong tool. That is a section of misaligned words, and the fixes in the previous section apply to that section.

It is on time and still reads late

Now the one you cannot fix by moving anything, because nothing is wrong.

A visual event and a sound register as simultaneous only when the visual arrives slightly first. Everyone learned this from watching music played: the stick is visibly moving before the snare sounds. So a word highlight that fires at the exact millisecond the vocal onset was measured reads as lagging, and beat-driven motion fired exactly on the beat reads as chasing the beat.

The renderer handles this with three leads, each measured against a musical event and each capped so it cannot overrun the previous one in a fast passage:

Lead	Value	Applies to
Word highlight	50 ms early, capped at half the gap to the previous word	The active-word highlight or wipe
Line entrance	Up to 80 ms early, scaled to the gap before the line	A line appearing before its first word
Beat anticipation	3 frames, about 100 ms at 30 fps	Beat-driven pulses and motion

There is also a release rule: the highlight normally holds until the next word starts, but if the next word is more than 250 ms away it releases at the word's end instead of lingering across a silence.

Until September 2026 there was a fourth number. The highlight led by 70 ms and the whole clock was then shifted 80 ms the other way, on the theory that a listener recognises a word at its vowel rather than its first consonant. The theory is fine. The arithmetic was not: the two nearly cancelled, and the aligner's onsets at the time were themselves about 27 ms late on average, so the net effect was a highlight roughly 10 ms behind an onset that was already behind. The clock shift was removed, the aligner's median signed error is now 0 ms on full songs, and one lead of 50 ms is the whole story. The reasoning for the model behind this is in how beat sync actually works.

If words still read late to you after all of that, you are in the previous section: it is taste, and the offset slider is the control for it.

The order to check in

1. Read the transcript against the song. Fix the text, then Realign.
2. Look for Timing chips. Nudge or tap those words.
3. Scrub the song in four places. If the error is the same everywhere, set the Lyric offset once.
4. If it still reads late and every word is measurably on its onset, that is the 50 ms question, and it is a preference. The offset slider is the answer to it too.

The first step fixes more lyric videos than the other three together, which is why it is first and why it is the one most people skip.

One song upload can become seven things in Revori, and the most common mistake is picking the format by habit: the artist who always makes a lyric video, the producer who always posts a visualizer. The formats are not interchangeable. Each one answers a specific question about a specific post, and the right way to choose is to ask that question for each post you plan, not once for the song.

Here is the whole set in one place, then the question each one answers.

Format	What it is	Shape	Where it goes	Credits
Lyric video	Words timed to the vocal, over footage	9:16, 1:1 or 16:9, 1080p; 4K on Studio	TikTok, Reels, Shorts; Instagram feed; YouTube	50 per render, 100 for 4K
Music video	Footage or photos cut to the song, lyrics optional	16:9	YouTube	50
AutoCut	Footage cut on the beat, no words	9:16, 1:1 or 16:9	TikTok, Reels, Shorts	50
Visualizer	Audio-reactive motion over your own backdrop	9:16, 1:1 or 16:9	TikTok, Reels, YouTube	40
Spotify Canvas	Silent loop, 3 to 8 seconds	9:16, 1080x1920	Behind the track in the Spotify app	25, Starter and up
Lyric card	A still image carrying one line	1:1, 4:5, 9:16 or 16:9	Feed posts, stories, X	5
Cover art	The release image	Square, 3000x3000 for Spotify and Apple Music	The distributor, every profile	From 18

Analysis runs once per song at 1 credit per second of audio and is shared across all of them. Editing, previewing and saving are free on every plan. The credits above are for renders only.

Start from the post, not the song

The question that sorts every format is: what is the listener going to do with this post? Read along and sing back. Watch a picture move to a sound. Glance at something in the corner of an app while the track plays. Screenshot a line. Recognise the release in a grid of other releases.

Those are different jobs, and the same song needs a different format for each of them, usually cut from a different section.

A hook people will sing back: lyric video

The words are the reason to stop scrolling. A chorus, a line that carries the song, a bridge everyone quotes. The lyric video shows those words timed to the vocal so the viewer reads and sings at once, over footage that sets the mood without competing with the text.

Vertical 9:16 for TikTok, Reels and Shorts, from the 20 to 30 seconds that carry the hook. A 1:1 cut for the Instagram feed and a 16:9 cut for YouTube are separate renders rather than crops, because a crop moves the text out of the safe area. The order of work that saves the most time is in how to make a lyric video.

Not for: instrumentals, heavily processed vocals where the words are texture, or a section with no lyric worth reading. Those are visualizer territory.

The full song on YouTube: music video

YouTube is the one platform where a full-length, landscape video is the native format rather than a repost. The music video module is 16:9, built from your footage or a set of photos cut to the song's structure, with lyrics as an option rather than the point. It is the asset that goes on the channel, gets linked from the streaming profile and lives as long as the song does.

Not for: vertical short-form. Post the lyric video there instead of a crop of this.

Footage but no words: AutoCut

You have clips, the song has a beat, and the words are either absent or not the point. AutoCut places footage against the beat grid so every cut lands on a musical boundary, with no lyric layer at all. It is the lyric video's editor minus the lyrics, which is why it costs the same and exports the same shapes.

Not for: a section where the listener should be reading. If they should, that is the lyric video.

Nothing to show: visualizer

No footage, an instrumental, a producer tag, the sixteen bars before the vocal comes in. A visualizer draws the audio itself, frame by frame, over a backdrop you choose: the cover, a photo, a video clip, one colour, a gradient. It skips transcription entirely, so there is no transcript to correct. Five styles, each drawing the sound differently, are described in the visualizer guide.

Not for: a hook with words. Motion over a line people want to read makes the line harder to read.

In the Spotify app: Canvas

The short vertical loop that plays behind the track in the Spotify mobile app, replacing the static cover while the song plays. Three to eight seconds, silent, looping without a visible seam. It does not need the song upload at all: a Canvas is built from a clip trimmed to a loop or a generated shot, and the crossfade at the loop point is applied for you.

The one thing it cannot do is follow the music. Canvas is not tied to playback position, so a loop timed to the chorus will land somewhere else for a listener who skipped in from a playlist. The full requirements and the loop-point craft are in Spotify Canvas requirements.

Not for: anything with a beat you want hit. Also not available on the free plan; Canvas renders start at Starter.

A line for the feed: lyric card

One line of the song on a still image. It is the cheapest format by a distance, at 5 credits, and it does a job none of the video formats do: it gets screenshotted, reposted to stories, quoted, and it loads instantly in a feed that autoplays everything else. Four shapes: square for the feed, 4:5 portrait for the feed with more height, 9:16 for stories, 16:9 for X and headers.

Not for: replacing the lyric video. A card is the line; the video is the line arriving on time.

The release itself: cover art

The square image the release ships with, 3000x3000 for Spotify and Apple Music, plus a 1:1 and a 9:16 for profiles and stories. Generated from a prompt on one of three engines, priced by engine at 18, 60 or 90 credits per image, with the most expensive being the one to use once you have found the direction on the cheap one. It does not need a song upload.

Not for: anything that moves. It is the one still the platforms require.

A release week, in formats

The formats stop competing once they are placed in time.

Weeks before, the cover art, because the distributor wants it first.

On release day, at the moment the track goes live, the Spotify Canvas, since it appears in the app immediately. Then the lyric video of the hook, vertical, because it is the post that gets shared and sung.

In the days after, a lyric card of the line people are quoting, a second lyric video from a different section if the first one travelled, and a visualizer for the intro or the instrumental version.

When the channel needs it, the 16:9 music video, or AutoCut over footage from the shoot.

Every one of those comes from the same upload and the same analysis. What changes is the section of the song and the question the post is answering.

Most font advice for lyric videos is a genre lookup table. Genre matters, but it is the second filter. The first is that video type has to survive conditions print and web typography never face, and that constraint eliminates most of the field before genre gets a vote.

The four conditions type has to survive

It is read in under a second. A line is on screen for the length of a phrase. There is no re-reading, no scanning back. Legibility at a glance is worth more than legibility on inspection, and those are not the same property. Faces with distinctive overall word-shapes beat faces with beautifully drawn individual letters.

It is small. A phone held at arm's length. Whatever the canvas resolution says, the delivered experience is a few inches of glass. Counters, the enclosed spaces in a, e, o and g, are the first thing to close up, and once they close the letters stop being distinguishable from each other.

It sits over moving footage. Contrast is not fixed. A word that reads cleanly over a dark frame disappears over a bright one two seconds later. This is what kills thin weights: a hairline stroke over busy footage has nothing to hold it.

It goes through compression. Every platform re-encodes. Compression is hardest on high-frequency detail, which is exactly what thin strokes, fine serifs and tight letterspacing are. Type that looked crisp in your editor comes back softened.

Those four conditions add up to one rule that does most of the work: heavy, open, and large-x-height, or it does not survive the trip.

What that rules out

Thin and light weights, below roughly 500. They look sophisticated on a still and vanish in motion.

Delicate serifs with fine hairlines. Didone faces in particular. Bodoni is beautiful and useless here.

Tightly spaced geometrics. Futura and its descendants have small x-heights and closed counters relative to their cap height, so at video scale they read smaller than they measure.

Anything with a script or handwriting model, unless the whole video is built around it. Connected letterforms are slow to parse and slow is the one thing you cannot afford.

The list, by genre

These are the faces we ship as preset defaults, which means they have been through real renders over real footage rather than chosen off a specimen sheet.

Genre	Face	Weight	Why
Hip-hop, trap, drill	Anton	400	Extremely heavy at a single weight, tight vertical rhythm, enormous x-height. Holds at 3-word lines over anything
High-impact, any genre	Bebas Neue	400	Condensed caps, fits more per line than Anton without losing weight. Single weight by design
Pop, K-pop	Poppins	800	Geometric but with a generous x-height and circular counters that stay open. Reads friendly rather than aggressive
Electronic, house, techno	Space Grotesk	700	Strong idiosyncratic letterforms, mechanical without being cold. The quirks survive compression because they are structural, not fine
Cinematic, marquee	Playfair Display	700	A high-contrast serif that works only at its heaviest weight and at large size. Use it big or not at all
Indie, editorial, folk	A mixed-case grotesque	400	Quiet, sparse, 3 to 4 words a line. The restraint is the effect
Minimal, one word at a time	A neutral grotesque	400	When one word fills the frame, the face should get out of the way

The last two rows are deliberately generic because our own presets use licensed commercial faces there, and recommending a face you would have to buy is not useful advice for a first video. Inter, Work Sans and Archivo cover that ground for free and get you most of the way.

Oswald deserves a mention as a middle point between Anton and Bebas: condensed, multiple weights available, less extreme than either.

Weight matters more than family

If you take one thing from this: picking the right weight of an adequate face beats picking the wrong weight of a perfect one.

Most of the failures we see are not a mismatched family. They are Poppins at 400 where it needed 800, or Playfair at regular where it needed bold. The face is fine. It is set too light to survive the four conditions above.

Anton and Bebas Neue both ship a single weight, which is not a limitation here. It is why they are hard to misuse.

Uppercase or lowercase

Uppercase is faster to recognise as a shape and slower to read as words, because word-shape recognition relies on ascenders and descenders that caps do not have. That trade is correct for three-word lines and wrong for anything longer.

All-lowercase reads as deliberate and informal. It suits hyperpop, bedroom pop and indie electronic, and it looks like an accident in a trap video.

Mixed case is the right default for anything that wants to read as considered rather than loud.

One face per video

Two typefaces in a 30-second vertical clip reads as indecision, not as hierarchy. If you need emphasis, you have weight, size, colour and position available before you need a second family.

The exception is a genuinely different element, such as an artist name or a release date sitting in a corner. That is not part of the lyric, so it can be set separately.

The practical test

Set your line. Export a still. Look at it on your phone at arm's length, over the brightest frame in your footage.

If you read it without effort the face is fine, and if you find yourself working at it the type is set too light. That test has been more reliable for us than any amount of specimen comparison, because it reproduces all four conditions at once.

Most lyric video tutorials describe the software. This one describes the order, because the order is where the time actually goes. Two of these eight steps decide whether the result is any good, and they are the two that feel like admin.

1. Choose the section before you choose anything else

Decide which 20 to 30 seconds you are posting.

Vertical short-form is a clip format. A three-minute lyric video posted to TikTok is not a lyric video for TikTok, it is a YouTube asset in the wrong place. The section that earns a replay is the one a listener already recognises, which is almost always the hook and rarely the opening bars.

Doing this first means every later decision is scoped to half a minute of material rather than to a whole song, which is roughly an eightfold reduction in work.

2. Upload and let the analysis run

Upload the audio. Three things happen: the vocal gets separated from the mix, the beat grid gets detected across the whole track, and the lyrics get transcribed with a start time on every word.

These are genuinely separate systems answering separate questions, which is why a tool that only detects beats cannot place words. If you want the detail, it is in how beat sync actually works.

Analysis costs 1 credit per second of audio in Revori and runs once per song. Reuse the same track for a visualizer or a Canvas later and you do not pay for it again.

3. Correct the transcript. Before anything.

This is the first of the two steps that matter and the one most people click past.

Sung vocals are hard. Layered harmonies, ad-libs, heavy processing, a producer's reverb tail, and words that are half swallowed by design. Any system will make errors, and if you are working in a language it handles less well it will make more of them.

The reason to fix them now rather than later is structural. The text is the foundation. Line breaks are computed from it. Timing hangs off it. Layout, sizing and emphasis are all downstream. Correct a word after you have styled the video and you have invalidated a chunk of the work sitting on top of it.

Read it once against the song. It takes a minute on a 30-second section. If the whole transcript is wrong, paste your own lyric sheet instead of correcting word by word.

4. Decide the look before you pick footage

Typography and colour, in that order, before you open the clip library.

One typeface, chosen for the genre. Heavy condensed display for hip-hop. Geometric sans for electronic. Rounded humanist for pop. A serif with air for indie. There is a longer version with the reasoning in the font guide.

Then two or three related tones, drawn from the mood of the song rather than from a swatch grid. Desaturated blue-grey for melancholy. High-saturation complements for club. Amber and teal for heartbreak. Red on black for anger.

Deciding these before the footage is a constraint, and the constraint is the point. Picking clips first and then trying to make type work over them is how you end up with a video where every element is fine and nothing agrees.

5. Place footage against the beat grid

Now the clips. The rule is that clip edges land on beats, ideally downbeats, so that cuts happen on musical boundaries instead of at arbitrary timestamps.

This is what the beat grid is genuinely for. It is not used to move the words, which stay pinned to the vocal performance where the singer actually put them.

One clip track is enough. If a section needs two things happening at once, it usually needs one better thing.

6. Cut the motion back

The second step that decides the result.

Motion is a budget. Every animated word spends some of it. Animate every word and it is all spent, and the viewer's eye has nothing to land on, because everything moving is indistinguishable from nothing moving.

Two or three large motion moments across a 30-second clip. Put them where the song has structure: the chorus entry, a key word, a section turning over. Between them, let the text sit still and let the music carry the rhythm.

If your tool makes it one click to give every word a different effect, that click is working against you.

7. Preview at phone size

Look at it in a phone-shaped viewport before you export.

Text that reads comfortably on a 27-inch monitor is routinely too small in a feed, and the failure is invisible at desktop scale. Three to six words on screen at a time is the working range for vertical. Anything denser gets skimmed instead of read.

Check the safe areas too. Platform chrome, captions and the interaction rail all cover parts of the frame that look empty in your editor.

8. Export at the right ratio

9:16 at 1080p for TikTok, Reels and Shorts. 1:1 for an Instagram feed post. 16:9 for YouTube proper.

Render each ratio you need as its own export rather than cropping one into another. A crop moves text out of the safe area and generally clips it. In Revori each ratio is a full separate render at 50 credits, because it is a full separate render.

Where the time actually goes

Step	Feels like	Actually costs
Choosing the section	30 seconds	30 seconds, and saves hours
Analysis	Waiting	A few minutes, unattended
Correcting the transcript	Admin	One minute, and decides everything after it
Type and colour	The fun part	Ten minutes if you decide before you browse
Footage	The fun part	Expands to fill whatever time you allow
Cutting motion back	Undoing your own work	Five minutes, and the largest single quality lever
Phone preview	Optional	One minute, catches the most common failure
Export	Done	Minutes, unattended

The pattern is that the two cheapest steps in wall-clock time are the two that determine the outcome, and both of them feel like they are getting in the way of the work rather than being the work.

Canvas is the short vertical video that plays behind a track in the Spotify mobile app. The format itself is simple enough to state in one line, and most articles about it stop there. The interesting problem is not the spec, it is the loop.

The spec

Property	Value
Aspect ratio	9:16, vertical
Delivery resolution	1080x1920
Duration	3 to 8 seconds
Audio	None. Any audio track is discarded
Playback	Repeats continuously for the length of the track
Content	No explicit imagery, no trademarks you do not own, no URLs, handles or calls to action in frame

Spotify for Artists is the authority on the content rules and on anything about upload and eligibility, and those do change. Treat the table as the shape of the format rather than as a substitute for checking before a release.

The two constraints that actually shape the work are the ones people skip past: it is silent, and it never stops.

Silent changes what you can do

You cannot hit a beat. Not because the tooling will not let you, but because it would not mean anything: a Canvas is not synchronised to playback position, so a listener who starts the track from a playlist sees the loop at an arbitrary offset. Any motion timed to a musical moment will land on a different moment for every listener.

So the visual has to work at any phase. That rules out the whole vocabulary of the beat-synced lyric video, which is a different format solving a different problem. A Canvas wants continuous motion, not punctuated motion.

Looping is the entire craft problem

A Canvas repeats with no gap and no fade to black. On a four-minute track, a 4-second loop plays about sixty times. The listener is not watching it closely, but they are watching it in peripheral vision for four minutes, and the join is the thing they will notice.

Three rules cover most of it.

One continuous shot, no cuts. A cut inside a 5-second loop reads as an edit, and once there is one edit in there the loop point becomes just another edit, which is worse. Keep it to a single move.

Motion that returns to where it started. A slow orbit, a rotation, a drift that comes back, smoke or water or light that has no fixed position to return to. What fails is directional motion: a push-in that ends closer than it began has nowhere to go but a hard cut back out.

Cross-fade the tail into the head. Even a well-chosen shot rarely ends exactly where it started. Overlapping the final second or two back over the opening frames hides the residual difference. We cap that overlap at two seconds, which is generous on a 3-second loop and about right on an 8-second one. Past that you are fading more than you are showing.

What to actually put in it

The honest answer is that a Canvas is atmosphere, not information. It sits behind the now-playing screen while someone listens. It is not a place to explain anything, and text in it competes with the track title and artist name that Spotify draws over the top.

What works: a single subject with slow ambient motion. Texture, light, water, fabric, smoke. A crop of the cover art with life in it. The artist, in one continuous move, doing very little.

What does not: anything with a beginning and an end, dense text, rapid cuts, or a visual that only makes sense the first time you see it.

Making one without shooting one

If you have the footage, any editor that can export 9:16 at 1080x1920 with the audio stripped will do the job. The work is in the choosing and the loop point, not the export.

If you do not, the two routes are a library clip trimmed to a loop, or a generated shot. Both are in Revori: pick a start point in a clip, set a loop length between 3 and 8 seconds, and the crossfade is applied for you, or generate a one-shot vertical loop from a prompt. A Canvas render costs 25 credits. The song analysis you already paid for on a lyric video is reused, so a Canvas from a track you have already uploaded costs nothing beyond the render.

The short version

Vertical, 1080x1920, 3 to 8 seconds, silent, looping. After that, the join is where the remaining effort should go, because it is the part a listener sees sixty times in one song.

Every tool in this category implies that you upload a song and get correct, timed lyrics. It is worth stating plainly what the actual ceiling is, because the gap between the claim and the reality is where people lose an afternoon.

The numbers

Measured on JamendoLyrics, which is 20 English songs with word onsets labelled by hand and released under MIT:

Metric	Our aligner	What it means
Mean absolute error	97ms	Average distance from the human-labelled word start
Median error	47ms	Half of all words land inside this
Within 50ms	52%	Roughly a frame and a half at 30fps
Within 300ms	95%	The tolerance the field generally treats as correct

For context, published systems as of 2024 sat around 200ms mean error. The bulk of a song lands tight.

Now the caveat that matters more than any of those numbers: the benchmark hands the aligner the correct lyrics. It measures alignment in isolation, deliberately, because that is the only way to test an alignment change without transcription noise swamping the result. It proves the aligner. It does not prove the product.

On a real upload, the text comes from speech recognition, and a wrong word is a timing error by construction. There is no timestamp for a word that was never sung.

Why sung vocals are hard

Speech recognition is built on assumptions that singing deliberately violates.

Pitch is supposed to be incidental to meaning. In singing it is the point, and it moves across a range speech never uses. Vowels are supposed to last a certain duration. In singing they are held for whole bars. Consonants are supposed to be crisp boundary markers. Singers soften them constantly because a hard consonant interrupts tone.

On top of that, the vocal shares frequency space with everything else in the mix. We separate the vocal stem before doing anything else, which helps a great deal and is not free: separation itself introduces artefacts, and it is imperfect on dense material.

The five things that actually break it

Layered harmony. Three vocal takes stacked in thirds are three simultaneously valid transcriptions of the same line, at slightly different times. The model has to pick one, and there is no principled basis for the choice.

Heavy processing. Autotune, doubling, heavy reverb and delay all smear the acoustic boundaries the aligner reads. A long reverb tail in particular makes a word appear to end considerably later than it was sung.

Ad-libs over the lead. A background vocal saying something different underneath the main line is, acoustically, another voice in the same stem. Ad-libs get interleaved into the main transcript in the wrong place more often than any other single failure.

Quiet and sparse passages. An intro sung against near-silence gives energy-based detection almost nothing to work with. Counter-intuitively these are often harder than dense sections rather than easier.

Languages outside the dominant training data. Accuracy tracks training data availability closely. Code-switching mid-line is the worst case, because the switch has to be detected before it can be transcribed across.

The tail

Averages hide the shape of the failure, and the shape matters more than the average.

Our remaining error is not spread evenly. Most songs sit near the median. A small number fail badly, mis-seating entire phrases rather than drifting by a few tens of milliseconds. On our benchmark the worst track has a mean error of 466ms and a 90th percentile over 1.3 seconds while the median song is fine.

This is why an average is a misleading way to describe a system like this. Ninety-five percent of words within 300ms and one song in twenty that needs real intervention are both true at the same time, and the second one is what a user experiences on the day it happens to them.

What honest looks like in a product

Given all of the above, the design question is not how to reach perfection. It is what to do about the cases that will not reach it.

Our answers, in order of how much work they save:

Show the transcript before anything is built on it. The review step exists because a correction is cheap before styling and expensive after. Every downstream decision, from line breaks to layout, sits on the text.

Accept a pasted lyric sheet. If you have the real lyrics, giving them to the system removes the transcription problem entirely and leaves only the alignment problem, which is the one with the 97ms number attached to it. It is the single most useful thing a user can do here, and it is why the feature exists.

Make re-matching one action. Sometimes the wrong song was matched. Correcting that line by line is absurd, so it is one button.

Keep manual editing free and unlimited. Retiming a line, nudging a word, or shifting the whole track with a global offset costs nothing and is not rationed. A system with a known error tail cannot also put its correction tools behind a paywall.

What we would tell you before you upload

If your track is a single lead vocal, moderately processed, in a well-supported language, expect the transcript to need a couple of corrections and the timing to be publishable as-is.

If it is dense, layered, heavily processed, or in a smaller language, expect to paste your lyrics and expect to nudge a section or two.

Either way, listen to it once before exporting. We keep an internal rule that the objective benchmark is for iterating and the human ear is the final gate, and it applies to your track for the same reason it applies to ours: the metric cannot hear.

For about six months, Revori's lyric alignment ended with a stage called onset snapping. Every piece of research we read described it as the load-bearing part: the thing that turns approximate word timings into tight ones.

We measured it against human-labelled ground truth. It was making the timing worse. We deleted it, and the median word landed 31% closer to where it belongs.

This is what we did, what the numbers were, and why we couldn't see it for six months.

What the stage did

Forced alignment gives you a timestamp for every word. Ours comes from wav2vec2 running on a demucs-separated vocal stem. The timestamps are good but slightly late, because an aligner tends to mark where a phoneme is clearly present rather than where it began.

The standard fix is to snap. You run a spectral onset detector (superflux, in our case) over the vocal stem, get a list of acoustic events, and move each word onset to the nearest one within some window. Ours used plus or minus 120ms. The logic is clean: the aligner knows which word, the onset detector knows when something started, so combine them.

Everyone does this. It is in the papers. It was in our own architecture doc, described in those words: the load-bearing lever.

Why we couldn't tell it was broken

Because lyric timing is judged by ear, and the ear is a terrible A/B instrument.

When you watch a lyric video, you are not measuring milliseconds. You are forming an overall impression, and that impression is polluted by the font, the animation curve, the footage cut, the loudness of the kick, and whether you already know what the next word is. Change the snapping and re-watch, and you will have an opinion. It will not be reliable. We had opinions in both directions for months.

Worse, the failure was not uniform. Snapping helped some words and hurt others. Averaged across a song, it felt like noise. Averaged across a catalogue with a metric, it was not noise at all.

So the actual fix was not an algorithm. It was building something that could tell us we were wrong.

The benchmark

We used JamendoLyrics: 20 English songs, 1520 words, with manually labelled word onsets, released under MIT. Human ground truth, not another model's output.

Two design decisions made it useful rather than merely existent.

Score a window, not a whole song. Real uploads are roughly 30 second clips, not four minute masters. Scoring full songs adds a long tail of instrumental traversal that no user ever sees, and runs about eight times slower. We take the densest singing window of about 35 seconds per song, which is both faster and closer to what production actually does.

Split prep from eval. The expensive part (demucs, then forced alignment) does not change when you vary the post-processing. So we run it once per song and cache the raw pre-snap alignment. Each variant then re-applies only its own post-processing step against that identical cached alignment.

That second decision is what makes the comparison trustworthy. Every variant scores against a byte-identical alignment, so the only thing varying is the step under test. It is a perfectly controlled A/B rather than two separate pipeline runs that happen to differ in one setting.

We report mean absolute onset error (AAE), median error, and PCO@50: the percentage of words landing within 50ms of truth, which is roughly a frame and a half at 30fps.

The result

mode	what it does	AAE	median	PCO@50
snap (legacy)	snap onset to nearest superflux onset, ±120ms	113ms	68ms	35%
off	trust the debiased aligner onset	98ms	48ms	51%
refine	nudge onset to the foot of its own energy rise, ±60ms	97ms	47ms	52%

Doing nothing beat the load-bearing lever. Turning snapping off dropped median error from 68ms to 48ms and took words landing within 50ms from 35% to 51%.

That is the whole finding. The most sophisticated stage in the chain was worse than deleting it.

Why snapping lost

Superflux onsets are not word onsets. They fire on any spectral flux increase, which in a sung vocal means vibrato peaks, consonant mid-points, breath, and the attack of a note that started three syllables ago.

So when you snap an already-decent onset to the nearest detected event, you move it away from truth roughly as often as toward it. The wins and losses cancel, and the losses are uglier, because a word yanked 90ms onto a vibrato peak reads as broken in a way that a word 40ms late does not.

We also tested the obvious response, which is to loosen detection so there are more onsets to choose from. That made it worse. More candidates means a closer wrong answer is always available.

What replaced it

The version we ship is called refine, and it is deliberately timid.

It anchors at the aligner's own onset and walks backward to the foot of that word's energy rise, bounded to plus or minus 60ms. It cannot jump to a different transient, because it never leaves the neighbourhood of the onset it started from. It only sharpens a leading edge that is already in the right place.

It also halves the systematic lateness. Mean bias went from +63ms to +24ms, which matters because a consistent lateness is exactly what reads as "not quite on the beat" even when nothing is obviously wrong.

The gain over simply switching snapping off is small: 48ms to 47ms median, 51% to 52% within tolerance. We kept it because the bias improvement is real and it costs almost nothing. But the honest summary is that most of the win came from deleting a stage, not from adding a better one.

It ships behind LYRIC_ONSET_REFINE, which takes refine, off, or snap, so the old behaviour is one environment variable away if we are wrong again.

What this benchmark does not prove

It uses perfect lyrics. Ground truth text is handed to the aligner, so the score isolates alignment quality from transcription quality.

Real uploads do not work that way. They depend on our ASR ensemble and the fusion step that reconciles it, and errors there produce timing failures that this benchmark cannot see. So the benchmark proves the aligner. It does not prove the product. A click track and a person listening remain the final gate before anything ships.

There is also a tail we have not fixed. Most of the remaining error sits in a handful of outlier songs rather than spread evenly. One track in the set sits at 466ms mean error with a p90 of 1309ms while the median song is fine. That tail is a different problem from the one this post is about, and it is the next thing worth paying for.

If you are building something similar

Three things we would tell ourselves six months earlier.

A stage everyone agrees is important is exactly the one to measure first. Consensus is why nobody checks. Ours had been described as load-bearing in our own documentation, which made it the last thing we suspected.

Build the controlled comparison before the improvement. The prep/eval cache took an afternoon. It is the only reason we could run seven variants and believe the ordering. Every hour spent on it bought back weeks of arguing about whether a change felt better.

Prefer bounded corrections to unbounded ones. Snapping could move a word anywhere within 120ms. Refine can only sharpen an edge within 60ms of where it already was. When your correction step cannot verify its own target, the bound is the thing keeping it honest.

The benchmark harness lives in scripts/bench/ and the ground truth data is a fetch script away, so any future alignment change has to beat 97 / 47 / 52% before it ships.

Revori is at revori.app if you want to hear what this sounds like on a real track.

The first lyric video I made took four hours.

It was a Sunday. I had a song I liked and footage I liked and wanted to put them together. I spent four hours in Premiere nudging keyframes a frame at a time, exported, watched it back, and the second verse had already drifted. Start over.

There was no profound lesson in it. I was annoyed enough to build the thing instead.

What the existing tools each get wrong

Three categories, three different failures.

Professional editors. Premiere, DaVinci Resolve, Final Cut. Total control and zero help. They do not know what a beat is. They do not know what a lyric is. Every sync is manual, and manual sync across a three-minute song is several hundred decisions you have to get individually right and then keep right through every re-edit.

Consumer editors. CapCut, InShot, the native TikTok editor. Fast and genuinely good-looking, and they do have beat features. But those features read the audio waveform, not the vocal. They will cut your clip on the drum hits. They cannot put the word "heartbreak" on the syllable where it is actually sung, because nothing in the pipeline ever transcribed it.

Lyric-video SaaS. The category that shows up in YouTube pre-roll. In principle these understand both beats and lyrics. In practice the transcription is where they fall down, and transcription is the part everything else depends on: a lyric video built on a bad transcript is unusable no matter how good the presets are.

None of these is wrong for its own purpose. None of them was the thing I wanted.

The thing that made it buildable

Two capabilities got good enough at roughly the same time.

Speech recognition now returns word-level timestamps reliable enough to build on, particularly when you run several models and reconcile them rather than trusting one. Beat detection got a genuine step change with neural detectors: we run beat_this, from ISMIR 2024, with a multi-band librosa detector as the fallback path.

So the mechanical half of the job, the half that took me four hours, is now machine work. What is left is the half that was always the point: which typeface, which footage, which colour, where the video should hold still.

That split is the whole design. The AI does the part with a correct answer. The person does the part that does not have one.

What actually happens when you upload

Upload a song. The vocal gets separated from the mix, the beat grid gets detected, and the lyrics get transcribed with a timestamp on every word. You review the transcript before you build anything, because no transcription is perfect and the review step is cheaper than discovering an error at export.

Then you pick a preset, pick clips from a library of 1,100 or more royalty-free videos or upload your own, and export a 1080p MP4. Default aspect is 9:16 because the destination is usually TikTok or Reels, with 1:1 and 16:9 available.

Analysis runs once per song. Reuse that same song for a lyric video, a visualizer, cover art and a Spotify Canvas without paying for the analysis again.

What it costs, plainly

Editing, previewing and saving are all free, with unlimited projects on every tier including the free one.

Credits are spent on two things only: 1 credit per second of audio analysed, and 50 credits for a standard 1080p render. The free tier starts with 500 credits and does not ask for a card.

Free-tier exports carry a small "Made with Revori" mark in the bottom corner. Paid exports do not. I would rather say that here than have you find it after you have built something.

What Revori is not

Revori is not a platform, a social network or an ecosystem. It does not collaborate with your team, match you with session musicians, or distribute your track to 400 streaming services. Each of those is a different product, and trying to be all of them at once is the standard route to a tool that does nothing well.

It takes your audio and gives back something you would actually post.

If you have ever built a lyric video by hand, you know the specific frustration: the word looks like it lands on the beat, you play it back, and it is a frame off. You nudge it. Now it is a frame off the other way. Sometimes it is exactly on the beat and it still feels wrong.

That last case is the interesting one, and it is not a timing bug. It is a sign that you are treating one clock where there are two.

Beats and words are two different clocks

A song gives you two independent timelines, and they are produced by completely different machinery.

The beat grid is a property of the whole mix. A beat detector listens to the full track and returns a list of beat times, which of them are downbeats, and the tempo. Ours is beat_this, a neural detector from ISMIR 2024, with a multi-band librosa detector as fallback when the model cannot load. It knows nothing about lyrics. It would return the same grid for an instrumental.

Word timing is a property of the vocal performance. It comes from forced alignment: you hand an aligner the audio and the text, and it returns a start time for every word. Ours runs wav2vec2 over a vocal stem that demucs has separated from the mix, because aligning against a full mix means aligning against a drum kit. It knows nothing about tempo. It would return the same timings if you halved the BPM of the backing track.

These two lists do not line up, and they are not supposed to. A singer who lands every syllable exactly on the grid sounds like a metronome. The push and drag against the beat is the performance.

What we do not do: move words onto the beat

This is the part most tools get wrong, and it is worth being blunt about because we shipped the wrong version of it for six months.

Every word we produce carries an annotation naming its nearest beat: nearest_beat_ms, the beat number, whether that beat is a downbeat, and a confidence score. That annotation is used to draw markers, to colour the timeline, and to decide where a clip edge should snap.

It is never used to move the word.

We did once run a snapping stage, though it snapped to detected vocal onsets rather than to beats. It was described in our own architecture doc as the load-bearing accuracy lever. When we finally measured it against human-labelled ground truth it was making timing worse, and deleting it dropped median word error from 68ms to 48ms. The full numbers are in the benchmark writeup.

The general lesson survives the specific case. A correction step that cannot verify its own target will move things away from truth about as often as toward it.

Where the beat grid is genuinely useful

Having said all that, the grid is not decoration. It carries four jobs:

Clip edges snap to it, so cuts land on musical boundaries instead of arbitrary timestamps. Section boundaries are labelled from it, so a chorus can look different from a verse. Beat-driven motion, the pulse or scale bump that makes a video feel like it is moving with the song, is triggered from it. And it gives you the visible ruler that makes manual editing tractable at all.

None of those require touching a single word timestamp.

Why exactly on the beat reads as late

Here is the effect that catches everyone.

A visual event and an audio event register as simultaneous only when the visual arrives slightly first. Live performance has physical latency baked into it. A drummer's stick is visibly moving before the snare sounds. A guitarist's pick crosses the string before the note. Watching music has trained everyone's expectations around that lead.

So an animation fired at the exact beat timestamp reads as lagging, even though the arithmetic is perfect. The fix is to fire early:

const { fps } = useVideoConfig();           // never hardcode 30
const BEAT_ANTICIPATION_FRAMES = 3;         // ~100ms at 30fps

const beatFrame = Math.round((beatTimeMs / 1000) * fps);
const triggerFrame = beatFrame - BEAT_ANTICIPATION_FRAMES;

Three frames is the value we ship. Below two the lead stops being perceptible. Past five the motion detaches and starts reading as early, which is a different kind of wrong.

The lead that used to be cancelled out

The anticipation offset is well known. This one is a correction of our own, and it is worth writing down because the mistake is easy to make.

Alignment pins a word to its acoustic onset: the consonant attack, or the start of an energy rise. That is the correct answer to the question the aligner was asked. A common theory says it is not where a listener hears the word, because word recognition anchors on the vowel, which for a consonant-initial word lands 50 to 100ms after the acoustic start. Until September 2026 we acted on that theory: the highlight led the onset by 70ms, and the whole karaoke clock was then shifted 80ms the other way to land on the vowel.

Measured, the two nearly cancelled. Worse, the aligner's onsets at the time were themselves about 27ms late on average, so the net result was a highlight roughly 10ms behind an onset that was already behind. The clock shift was removed. The aligner now takes a second opinion from other timing estimators on every word, its median signed error is 0ms on full songs, and there is one visual lead: the highlight fires 50ms before the onset, capped at half the gap to the previous word so a fast run stays in order.

Reading lags listening, and visual-early is the forgiving direction. Visual-late reads as off almost immediately. That asymmetry is the whole reason the lead exists and the reason it is small.

The three leads, in one place

Lead	Value	What it applies to	Why
Beat anticipation	3 frames, ~100ms at 30fps	Beat-driven pulses and motion	Visual must lead audio to read as simultaneous
Word highlight	50ms, capped at half the gap to the previous word	The active-word highlight	Reading lags listening; early is the forgiving direction
Line entrance	Up to 80ms, scaled to the gap	A line appearing before its first word	The eye needs to find the line before it needs to read it

All three are measured against a musical event, and every one is capped relative to the surrounding gap, so none can overrun the previous word during fast passages. A fixed 80ms lead is fine at 90 BPM and catastrophic in a triplet run. There is also a release rule: the highlight holds to the next word unless the next word is more than 250ms away, in which case it lets go at the word's end instead of lingering across the silence.

If the result still reads early or late to you, the Sync card in the Style panel has a per-project offset from 2,000ms early to 2,000ms late. The four different reasons lyrics can look off, and which tool fixes each, are in why the words look early or late.

What this adds up to

Beat sync sounds like one feature and is really a stack of small decisions, most of which are about perception rather than arithmetic. You detect the beats, separate the vocal, align the words against the vocal and leave them there, use the grid for cuts and motion, then apply three small leads, each bounded so it cannot run into its neighbour.

None of that is visible in the finished video. It shows up only as words landing where you expect them to.

Most lyric videos on TikTok are bad, and it is usually not a taste problem. It is a tooling problem. The tools make the decorative choices easy and the structural ones invisible, so people spend their effort on the layer that matters least.

Four things actually decide whether a lyric video holds someone. None of them is an effect.

1. Does the typography match the genre?

Typography is a genre signal before it is a legibility decision. A face carries a set of associations, and using the wrong one tells a viewer you are outside the music, which is a harder problem than looking plain.

The rough map, which is also roughly what our own presets ship with:

Heavy condensed display faces for hip-hop and trap. Anton, Bebas Neue, Oswald. The visual weight matches the low end.

Geometric sans with strong letterforms for electronic and house. Space Grotesk, Montserrat at its heaviest. Clean, synthetic, repeating.

Rounded humanist sans for pop. Poppins, Inter, Nunito. Approachable, which is the register pop is in.

A serif with air for indie and folk. Playfair Display, Cormorant Garamond, Instrument Serif. It should suggest a page rather than a billboard.

Italic scripts or oblique serifs for R&B and soul, where the curves echo the melisma.

There is a longer version of this with the reasoning per genre in the font guide.

2. Does the motion arrive before the beat?

If your animation fires on the exact beat timestamp, it will feel late. This is not a preference. Visual events register as simultaneous with audio only when the visual arrives slightly first, which is why a drummer's stick is visibly moving before the snare sounds.

The practical number is about 100ms, or three frames at 30fps. That single offset is the difference between motion that feels locked to the track and motion that feels like it is chasing it, and it is invisible until you flip it back and forth.

The word highlight has its own version of the same rule: it fires 50ms before the vocal onset, because reading lags listening and early is the forgiving direction, while a highlight that arrives late reads as off almost immediately. Both leads are covered in how beat sync actually works.

3. Does the colour follow the mood, or fight it?

Pick the colour story first, before the font, before the footage. Everything downstream gets easier and the result is coherent rather than assembled.

Melancholy wants desaturated palettes. Pull the vibrance out, sit in blue-grey, keep text warm white and never hot.

Party and club want high-saturation contrast. Magenta against cyan, lime against navy. Colours that look lit by a fog machine, because they were.

Heartbreak sits in amber and teal, which is the cinematographer's default for a reason and reads as film stock rather than filter.

Anger wants pure red and pure black. No gradients, nothing else in the frame.

The fastest way to a video that feels wrong is to skip this step and pick the brightest options from a preset picker. Nothing about it will look broken. It will just feel like it belongs to a different song.

4. Do you know when to leave the screen still?

This is the one nobody talks about and it is the largest of the four.

Watch a music video you actually like and count the frames where the text is moving. It is a small fraction of the runtime. Most of the video is type sitting still while the rhythm happens in the music instead of on the screen.

Motion is a budget. Every animated word spends some of it. Animate every word and you have spent all of it, and the viewer's eye has nothing to catch on, because everything moving is the same as nothing moving.

Good lyric videos have two or three large motion moments per song, placed where the song already has structure: a chorus arriving, a key word, a section turning over. In between, the text sits still.

A tool that makes it easy to give every word a different effect is working against you. Our defaults animate the first word of a line and then hold until the next line starts. You can override that, but you have to decide to, which is the point.

The four, side by side

Element	Get it right	Get it wrong
Typography	One genre-appropriate face, used throughout	A face chosen for novelty, or several per video
Motion timing	Fires ~100ms before the beat	Fires exactly on the beat, reads as late
Colour	Palette picked first, from the mood	Brightest preset, picked last
Restraint	Two or three motion moments per song	An effect on every word

The short version

Good lyric videos are correct more than they are fancy: genre-appropriate type, motion that anticipates the beat, colour that matches the mood, and restraint. Getting those four right matters more than any effect you could add on top, and getting one of them wrong is not something an effect will cover.