How to make a lyric video

Putting words on screen is the easy half. Putting them on screen at the moment they are sung is the half that decides whether it looks finished — and it is the half every automatic tool gets wrong. This is the method, including the parts you can do in any editor.

Why automatic timing fails on singing

Every tool that offers to make a lyric video for you starts by transcribing the song. That step is doing something it was never built for. Speech recognition is trained on speech, and singing breaks its assumptions: vowels are held for whole bars, single syllables are spread across several notes, and the voice sits under a band occupying the same frequencies.

There is a specific failure worth knowing about, because it explains results that otherwise look inexplicable. Most transcribers runvoice activity detection first, to decide which parts of the audio contain speech at all. On a full mix that step regularly decides a sung line is not speech and throws it away before anything has tried to read it. Testing this on a three-and-a-half minute song, a transcriber matched 1.5% of the lyric with that filter on. The same song, same model, filter off:93%.

They also fail loudly rather than quietly. A transcriber that loses sync does not stop — it emits words with timestamps piled on top of each other. Six words inside 0.12 seconds is fifty words a second, which nobody sings, and any timing built on it puts a whole verse in the wrong place.

Type the words yourself

The fix is not a better transcriber. It is to stop asking the machine what was sung and ask it only when.

You have the lyrics. They are correct, they spell the names right, and they contain the words the singer slurred. Supply them as text and the only remaining question is timing — which is a much easier question, and one where being slightly wrong is survivable. A mistimed line looks approximate. A wrong word looks broken, especially to the people most likely to watch, who know the song.

Break the lines the way the song is phrased. That is not cosmetic: the line break decides what appears on screen together, and a lyric arriving as one long paragraph has to be chopped up by a word count instead, which lands mid-phrase and reads as a transcript with a beat.

Timing the lines

Work in two passes. Let something place the lines automatically first — matching your typed words against what it hears will get most of a song close, and it saves the bulk of the labour. Then correct it, because it will be wrong somewhere and it is always wrong in the same places: where the vocal is buried, and where the song repeats a line it has sung before.

The fastest correction is to tap the lines in. Play the song and press a key at the start of each line. A lyric video is about fifty lines, so a full pass takes roughly the length of the song, and tapping is more accurate than it sounds because you are listening rather than staring at a waveform.

Two things about tapping that are worth knowing before you blame yourself. You will always tap late: reaction to an anticipated sound is about 150 milliseconds, and your computer's audio output adds 100 to 200 more, so every tap lands roughly a third of a second after the beat. Good tools subtract that. And the timing must be judged against the audio clock rather than a general one — those two drift apart, and a tap compared against the wrong clock is compared against a song playing somewhere else.

What to show

Most people making a lyric video have a song and a piece of cover art, and no footage at all. That is enough, if the still moves.

Put a slow push or drift on the artwork — a Ken Burns move. A photograph that sits perfectly still reads as a stalled video, while the same image drifting slowly reads as a deliberate shot. Between sections, color fields taken from the artwork's own palette keep everything looking like it belongs to the same record, which stock gradients do not.

Whatever you choose, keep it quiet. The words are the subject. Anything with detail or fast motion behind them competes for the same attention and wins, which is the most common way a lyric video made with real footage ends up harder to read than one made with a single photograph.

Making the words readable

• A line at a time, not a word. One word alone gives the reader no context. Show the line and highlight the word being sung as it passes.
• Outline the type, don't just shadow it. A white caption over a bright background disappears. A dark outline survives anything underneath it.
• Sweep the highlight, don't jump it. If the highlight moves word to word, timing that is half a word out lights the wrong word and reads as broken. A fill that wipes across reads as slightly early or late, which is what karaoke has always looked like.
• Never let a line flash. If words are on screen for a tenth of a second each, nobody reads them. Give a line the time its words need even when that means leaving a gap afterwards.

Exporting

1080p at 30fps, H.264, around 7 to 8 Mbps. That is YouTube's own recommendation for 1080p30, and it matters more here than for camera footage because smooth color backgrounds are the hardest thing there is to encode cleanly. Drop the bitrate to save space and banding appears in exactly the gradients you chose for their smoothness. A three-minute video will land somewhere near 150 to 180MB, which is normal.

Put your logo bottom left. YouTube stamps the channel logo in the bottom right of every video, so anything there ends up underneath it.

Then put the full lyrics in the video description. A search engine cannot read the words burned into your video, so the description is the only text it has — and it is what makes a lyric video findable by the line somebody half-remembers.

Rights, briefly

If you did not write the song, a lyric video is a derivative work and you need permission from whoever holds the rights. "It is promotional" is not a license. If it is your own song, credit the writers and the performer in the description anyway — it is what makes the video useful to anyone who finds it.