Simple tips for writing better AI song lyrics

I’ve been generating music with suno.ai for almost a year now. The v3 model that dropped last week is a huge step up—crossing the line from “curiosity” to “OMG, I’ve had an AI-generated song stuck in my head all day.”

I’ve actually subscribed for the paid account—which is insane, since the main way I use the app is absolutely trivial: songifying movie monologues and family inside jokes with my pre-teen kids. Yes, I am paying $10 per month to generate up to 500 2-minute songs.

Anyway, I’ve learned a lot in the process. This blog post covers some simple tips for writing much better lyrics. Enjoy!

1. Find lyrics, don’t generate them

The fastest, easiest, and probably worst way to create lyrics is through the Song Description field. The AI-generated lyrics are uninspired, and the rhyming is forced¹. Every once in a while, you’ll get a turn of phrase that feels like it could be part of an actual song, but that’s rare.

I prefer to switch into “Custom mode” and paste in my own lyrics. Suno does a remarkably good job of taking text—almost any text—and setting it to music.

• For example, here’s an acoustic gospel song I generated directly from the text of a Wired article titled “Why the East Coast Earthquake covered so much ground.” It’s fascinating hearing Suno find the natural cadence in sprawling sentences like “The United States Geological Survey is urging the region to prepare for aftershocks of smaller magnitude.”

• Another example: this “sad girl piano ballad” cover of the MIT open source license went viral on twitter last week. I promise that you’ve never heard a more entrancing legal document.

Directly copy-pasting works best if you pick a Style of Music that focuses on vocals and emotional depth over beat and meter: jazz, folk, soul, blues, gospel and R&B. (Opera is good at this, too, but my kids hate it.) Adding words like “singer-songwriter,” “ballad,” “expressive melody,” “expressive phrasing,” “lyrical,” and “acoustic” to the prompt seems to help as well, although it’s hard to be really sure which cues Suno is picking up on.

You can find interesting lyrics almost anywhere: a passage in a book, an error message, whatever’s at the top of your news/social media feed. My kids and I get a lot of mileage from songifying things that we quote a lot: movies, catch phrases, other songs, jokes. Putting these things to music instantly makes them more interesting, giving them new dimensions and life.

Found lyrics are almost always more interesting than generated lyrics. They came from somewhere, bringing backstory, context, and meaning with them.

2. Add line breaks

Like poems, real songs almost never use paragraphs. They often break the normal rules of sentences and punctuation. Instead, they pay close attention to phrasing and (sometimes) rhyming².

Adding line breaks is a great first step to turn text into song lyrics, by emphasizing phrasing.

For example, suppose you found this note³ on your fridge:

This is just to say I have eaten the plums that were in the icebox and which you were probably saving for breakfast.

Forgive me. They were delicious. So sweet and so cold.

With a little work, that could turn into:

This Is Just To Say

I have eaten
the plums
that were in
the icebox

and which
you were probably
saving
for breakfast

Forgive me
they were delicious
so sweet
and so cold

I find that adding line breaks is more useful for me than for Suno. Suno doesn’t seem to need this. As you can see from the earthquake and MIT license examples above, it does a surprisingly good job pulling phrasing out of all kinds of text, whether or not it’s marked by line breaks.

It’s still worth doing, for two reasons. First, adding line breaks flips a switch in my brain. It helps me hear the text instead of just seeing it. Second, adding line breaks makes it easier to make lots of other kinds of revisions, including those described further on in this blog post.

Here’s a dreamy hip hop version of This is Just to Say, generated directly from the original text of the poem. For the rest of this post, I’ll use this poem as a running example.

The first 30 seconds of this track feel more like a poetry reading than a full-production song. The vibe and vocals are great, but the poem is over almost too soon. Suno seems to agree, because it peppers the rest of the track with repeated phrases “they were delicious,” “forgive me,” “so sweet and so cold.” This works reasonably well for the ambient, slow-moving vibe of dreamy hip hop, but for most other musical styles it would be too spare.

3. Repeat lines

One of the simplest things you can do to improve your song lyrics is repeat lines. It might feel lazy to re-use lyrics this way, but real songwriters do it all the time. Repeating a line makes sure that it gets emphasized. It can help you control the pace and timing of the song.

Depending on the style of music, Suno will often take advantage of repeated lines to add interesting transitions and vamps, or add emphasis with dynamics and harmony.

Here’s a second Dreamy hip hop version of This is Just to Say, with a few lines repeated. To my ear, it sounds much more complete as a song.

This is just to say,
This is just to say.

I have eaten
the plums
that were in
the icebox.
The icebox

and which
you were probably
saving
for breakfast
Saving for breakfast.

Forgive me
Forgive me
they were delicious
so sweet
and so cold

So sweet
And so cold
And so sweet
And so cold

So cold.

3. Add backup vocals

If you wrap a line in parentheses, Suno will usually treat it as a backup vocal.

I have eaten
(Eaten)
the plums
that were in
the icebox
(The icebox)

This can be a great way to add contrasting voices. Also, it will often prompt Suno to do interesting things with dynamics and harmony.

Here’s a dancepop version of This is Just to Say. Suno has taken the parenthesized lines and converted them to a filtered echo of the main vocals. You can also hear how repeating lines works well with a highly rhythmic musical style like dancepop.

This is just to say,
This is just to say.

I have eaten
(Eaten)
the plums
that were in
the icebox
(The icebox)

and which
you were probably
saving
(saving)
for breakfast
Saving for breakfast.

Forgive me
Forgive me
they were delicious
(Delicious)
so sweet
and so cold

So sweet
And so cold
And so sweet
And so cold

So cold.
[Fade out]

Note: I’ve added “[Fade out]” at the end of this song, but I’m not sure it actually does anything. Suno doesn’t sing it as a lyric, but only some of the songs that include this cue actually fade out. In addition, I’ve experimented with other cues like [Quietly] and [Very fast], and Suno seems to ignore them completely.

4. Add vocalizations

Last, you can add vocalizations to your lyrics: “oh,” “aaaah,” “hey, hey, hey,” “ooooooo,” “na na na na,” and so on. Suno does an amazing job using vocalization cues to make songs more expressive.

Here’s a second dancepop version of This is Just to Say, this time with vocalizations. Suno uses the “ohs” and “aaaahs” to play with timing: lining the lyrics up with the phrasing of the music.

If you listen closely, you’ll notice that the vocalizations in the generated audio don’t line up perfectly with how I wrote them in the lyrics. Just adding a few vocalization cues in the lyrics seems to give Suno permission to make the song more flexible and expressive. If you want tight control, this could be a bad thing, but if you want that level of control, you’re probably not using Suno to begin with.

This is just to say,
This is just to say.

I have eaten
oh!
(Eaten)
the plums
that were in
the icebox
Oh!
(The icebox)

and which
you were probably
saving
(saving)
Aaaah
for breakfast
Saving for breakfast.

Forgive me
Forgive me
Oh, oh
they were delicious
(Delicious)
so sweet
and so cold

Aaaaah

So sweet
And so cold
And so sweet
And so cold

Oh, oh

So cold.
[Fade out]

Here’s a folk version of This is Just to Say. It uses exactly the same lyrics. The only difference is the Style of Music prompt: “folk” instead of “dancepop.” Again, Suno does an incredible job using the ohs and aaahs to make the song more expressive and more fun. I know that AI is just a network of weights predicting tokens, but the way it uses vocalization makes it sound like it’s having fun with the song, and is trying to share that fun with the audience. I love it.

One more example: Chef Skinner’s Lament, a song my 11-yo daughter “wrote” from the perspective of the villain in Ratatouille. We generated the lyrics using ChatGPT, with predictably awful results. That said, the instrumentation and vocals are delightful, and the flourishes around the “oh oh oh” and “ahhh…..” almost save the song.

Last thought on vocalizations: if you’re like me, putting these into your songs is faintly embarrassing, in a caught-singing-loudly-at-a-stoplight kind of way. It’s also impossible to imagine how they’ll actually sound, since you don’t know rhythm, meter, or melody. I’ve found that it’s best to not overthink it: just drop them in here and there and see what Suno comes up with. They almost always improve the song.

Putting it all together

Playing with Suno is delightful. It feels endlessly creative, and it’s fast: you can generate songs faster than you can even listen to them. Writing this blog post took my far more time than it took to generate the songs for it.

I have no idea how copyright law is going to evolve to deal, or what this technology is going to do to the music industry and other professions that use music. For now, I’m just having fun playing with a creative tool that’s unlike anything I’ve ever heard before.

I’ll wrap up this post with a few of my kids’ favorites from the last few days.

I Found Some Web Results, classic rock version

Backstory: we have an Apple HomePod in our kitchen. In theory, you can ask Siri any question, and it’ll look the answer up online and share the result. In practice, Siri is laughably bad at this.

On top of other misunderstandings and glitches, Siri is prone to refuse to answer on the homepod. Instead, it says “I found some web results. I can show them if you ask again from your iphone.” For some reason, my kids think this is hilarious, and mock Siri by trying to recite the message together every time the message comes up.

I Found Some Web Results (Gospel choir version)

Exact same song and lyrics, but this time sung by an AI gospel choir, a peppy call-and-response number that would have felt at home in Sister Act. This is now my ringtone.

Justice is a Non-Corrosive Metal (Broadway musical ensemble)

Sometime last year, my 13-yo memorized the “Microwave of evil” dialogue from Dreamwork’s Megamind. To this day, he will recite it at the drop of a hat.

Using Suno, we converted the dialogue into an absolutely unhinged ensemble number from an imaginary Broadway production. Don’t think too hard about which voices are singing which lines. Or why Megamind has a bad French accent. Or why the song is set to a faux-calypso beat.

We loved this song so much that we actually bothered to get Suno to finish it, using the somewhat glitchy “Continue From this Song” function. I like to think that this is what Sir Andrew Lloyd Weber would create if he had the time to produce Megamind, the musical.

1. Keep an eye out for “torn” and “bold.” ChatGPT uses these words to death.

2. Rhyming is hard. We think that “song lyrics should rhyme,” but that’s not really true. Except for a few artists and genres, most lines in most songs don’t rhyme. Instead, you usually see rhymes salted throughout songs, especially in a few key places that the songwriter wants to be memorable, such as the chorus, or the last stanza of an important phrase.

3. From William Carlos Williams’ poem This is just to say