My Adventure in AI Songwriting

 My lovely wife, Abigail Darwin, surprised me on Valentine’s Day with a charming ballad about our relationship entitled “From Darwin Day to Forever”. It was sensuous and beautiful, and I was blubbering like a baby before it was over. Of course I had to know how she had come up with it, and she said she had provided ChatGPT with a lot of the personal details and asked it to generate the lyrics to a love song, which it had done. She then fed those lyrics into OpenMusic.ai to have it gin up the melody, instrumentation, and vocalist.

Well, this inspired me to try it myself, and for a subject I chose pickleball, and for a rhythm I picked the polka, which has been the Wisconsin state dance since 1993. These are the instructions I fed into ChatGPT to get the lyrics:

= = = = = =

I want you to create lyrics to a parody song in polka timing entitled “The Pickle Barrel Polka”. Do it in the style of Frankie Yankovic and His Yanks doing “Beer Barrel Polka”, and please credit the song to Yankee Frankovic and Half a Dozen Chubby Belgians. The lyrics should rhyme. They’re intended as an homage to the sport of pickleball and so MUST include references to a paddle or paddles and “the kitchen” and SHOULD include references to SOME of these aspects of the game: starting with a score of 0-0-2, slams, dinking, keeping your eye on the ball, 3rd-shot drops, “play to 11 win by 2”, stacking, and aspirations of being like superstars Ben Johns and Anna Leigh Waters. FYI, the rating system in pickleball is DUPR, pronounced to rhyme with “super”. Since it’s for circulation in the area of Madison, Wisconsin, it needs have a reference to CAPA (Capital Area Pickleball Association), plus at least SOME of these outdoor courts — Garner Park in Madison, McGaw Park in Fitchburg, Taylor Park in Middleton, Bakken Park in Cottage Grove, McFarland Park in McFarland, Wyndham Hills Park in Sun Prairie, and Kilkenny Farms Park in Waunakee — and ALL of these indoor facilities — Dill Dinkers, Pickleball Kingdom, Pickle Pro Courts, and Hitters.

= = = = = =

I hoped that that wasn’t too much to ask, but the result was terrific: about 95% ideal, only requiring a few tweaks from me. Some of them involved either adding or subtracting syllables to achieve proper scansion, I changed the provided “0-0-2” to “oh-oh-two” to avoid a pronunciation of “zero”, more colorfully described the results of a popup as having your teammate “eat it”, followed a reference to “Anna Leigh and Ben” with “oh my!” instead of “on high”, replaced “neighbors” enjoying the game with “players”, and used “bestride it” (“it” being the kitchen line) rather than “within it” to rhyme with “inside it”. Then there was this odd couplet:

“Game’s to eleven, win by two!” the scorekeeper will cry,

Then somebody argues softly — but politely, mid-July!

That 2nd line made no sense at all, so I replaced it with something that was more score-related:

No game will last forever, nor ending in a tie.

Ready for prime time, I figured, and submitted it to OpenMusic.ai to jollify it with a tune.

= = = = = =

But alas, the lyrics were rejected as inappropriate, so I asked ChatGPT why, and this is what it said:

Good news: Nothing in your lyrics is actually offensive to a human.

Bad news: You accidentally tripped several classic automated-moderation landmines — the kind music generators are notoriously oversensitive about. These systems don’t understand context (sports humor). They scan for words statistically associated with violence, injury, substance use, or harassment. Your pickleball jokes accidentally look like ... assault and drinking songs.

It overconfidently said “These are almost certainly what got you rejected” and cited the input phrases  “eat it mighty soon”, “slams it”, “try to miss my face”, “whack-whack-whacks”, “guard the kitchen fiercely”, “shout”, and even “foot fault” as contributing to the “hostile argument” classification.

Adding to the negativity scoring were the alcohol references (“snacks and beer” and “cooler in the cooler”). ChatGPT acknowledged that “Your song sounds like a recreational sporting event — but the classifier sees substance promotion” and added that “Many music AIs block alcohol because they are also used for kid-safe generation modes.... Moderation is cumulative — multiple weak signals can cross the threshold.”

It advised “You only need to soften a few lines. Keep the jokes — remove injury wording” and specifically suggested changing “eat it mighty soon” to “scrambling mighty soon”, “try to miss my face” with “that was quite the pace”, “snacks and beer” with “snacks and cheer”, “guard the kitchen fiercely” with “guard the kitchen neatly”, and “cooler in the cooler” with “laughter in the cooler”.

= = = = = =

So I rolled my eyes and revised the “violent” parts to see if that would suffice (but left in the alcohol references). Nope. Still rejected. I tried again, this time additionally replacing “beer” with “bheer” and “a cooler in the cooler” with “a dip inside the cooler”. But this attempt at wussification was likewise ixnayed, again for reasons unknown and unspecified.

= = = = = =

At this point I started to think that maybe it was just this particular website that was so priggish, so I asked Google Gemini for recommendations for a free song-writing service that might be more willing to take whatever lyrics I gave it. These were the ones it recommended:

1. Suno.com (the most beginner-friendly) is currently the “gold standard” for turning text into full songs. It understands specific musical instructions better than most.

2. Udio.com is often cited as having slightly higher “studio-grade” sound quality, though its free tier is a bit tighter.

In both cases I was advised to start by selecting “Custom Mode”, then to paste my lyrics into the Lyrics box and, in the Style of Music box, enter “polka, upbeat, accordion, oom-pah rhythm, male/female vocals” before clicking “Create”. Additional tips were:

 • Use Brackets: AI models respond to structural cues. Add tags like [Verse 1], [Chorus], and [Accordion Solo] between your lyrics to tell the AI exactly when the “oom-pah” should kick into high gear.

 • Timing is Key: Polka is almost always in 2/4 or 4/4 time. If the first result feels too slow, add the words “Upbeat” or “High BPM” [apparently assuming that I knew it meant “beats per minute”] to your style prompt.

 • The “Polka” Tag: Don’t just say “polka.” Adding “German Polka” or “Czech Polka” will often trigger different instrumental palettes (more brass vs. more accordion).

= = = = = =

That latter point was new to me, so I inquired about polka styles and was informed that:

Czech, Polish, and German polka styles each have distinct rhythms and instrumentation:

 • Czech (Bohemian): Lively 2/4 rhythm, often with accordion, clarinet, trumpet, and bass. Melodies are crisp and dance-focused, rooted in folk traditions.

 • Polish: Stronger off-beat accents, with prominent accordion (especially piano accordion), trumpet, and drums. American Polish polka often includes saxophone and has a driving, festive energy.

 • German (“Dutchman”): Softer, smoother feel—common in Wisconsin and Midwest taverns. Features accordion, tuba, clarinet, and guitar, with a steady, swaying rhythm suited for group dancing.

Each reflects its cultural roots while evolving in American communities.

= = = = = =

Well, “Wisconsin taverns” settled me on German style. But I had a few additional questions as well: “Can I give either of these AIs instructions as to who should get song-writing credit? Or name a particular polka band whose style I’d like to imitate? Or would that just confuse it, and I should only provide the lyrics?”

Response:

When it comes to the “behind the scenes” instructions for these AIs, there are some very specific rules to follow if you want the best results without getting “blocked” by their safety filters. Here is how to handle credits and styles without confusing the system:

1. Can you give song-writing credit? Not directly in the prompt. These AI models don’t have a “metadata” field for credits that will show up on the file itself [but] If you want credits to be part of the song, you can literally include them as spoken lyrics at the beginning or end. (A Note on Ownership: In 2026, most platforms like Suno and Udio grant you commercial ownership of the song only if you are on a paid plan. On the free tier, the platform technically owns the output, though you are free to share it for personal use.)

2. Can you name a specific band? Generally, no. Both Suno and Udio have “Artist Filters” to avoid copyright lawsuits.... If you type “Write a song in the style of Frankie Yankovic,” the AI will likely flag the prompt and refuse to generate it, or it will automatically strip the name and replace it with generic tags like “Accordion Polka.”

3. Should you provide ONLY lyrics? Definitely not. If you only provide lyrics, the AI will guess the genre — and it almost never guesses “Polka” by default. (It usually defaults to Pop or Alt-Rock).

= = = = = =

What I decided to try first for style was “German style polka, 2/4 time, oom-pah, fast accordion, festive, cheerful, male and female vocalists”. And, since I’d been keeping separate text files for the different versions of the lyrics, I reverted to V.2, my initial revision of ChatGPT’s V.1 (except I retrofitted V.4’s “a dip inside the cooler” into V.2 because I actually liked it better anyway). And I fed that into Suno, which turned me down because it contained “inappropriate material”. So I tried again with the tamed-down lyrics of V.4 (skipping right over V.3, which omitted only the “violent” stuff while retaining the drinking references). Again, no go. Shoot!

Decided to try Udio. It asked me to select a genre. Obviously what I wanted was polka, which did not appear in their list of pop, electronic, hip-hop, R&B, rock, country, jazz, classical, reggaeton, or eclectic. Kissed that one off right away.

Back to Suno. I tried the old programmer’s trouble-shooting trick of sending it the text in halves, to see which half it choked on, then cutting that in half until the problem area was narrowed down. But both the 1st half and 2nd half got bounced. Dang, dang, shooty bang.

OK, try each 4-line stanza individually. And whoopee, Suno didn’t care about my teammate having to eat the ball, so Verse 1 was OK. Finally, a good result! Actually, it gave me two results, one half a minute long and another about 3 times that length, with an instrumental interlude between the two repetitions of the verse. (It subsequently turned out that it would always give me two renditions to choose from.) The specification of “fast accordion” was too fast, tho, so I omitted “fast” from the style spec. (Altho, as it turned out, the platform had an option to adjust speed, but I preferred to get it the way I wanted without undue subsequent tweaking.)

Onward:

 • Chorus rejected.

 • Verse 2 Stanza 1 OK. No problem with whacking. Based on the music’s speed, I decided to modify the style to say “not too fast” and drop the “festive”, which may have implied speed.

 • Verse 2 Stanza 2 OK. No problem with “slam it” or “try to miss my face.” “Not too fast” was still too fast, so I changed it to “moderate pace”.

 • Next chorus OK. No problem with “cooler” being unduly boozy, but the pronunciation of “DUPR” was off, so I added an “e” to make it “DUPeR”. And “average” was slurred into 2 syllables, so I changed it to “av-er-edge” to indicate syllable breaks.

 • Verse 3 Stanza OK. No problem with beer.

 • Verse 3 Stanza 2 OK. No problem with “fiercely” or “fault”. The vocalist in the long version went nuts throwing in lots of unscripted “yi yi yi”s; that one was doomed. But that’s what led me to add “danceable” to the style specs.

 • Last chorus rejected.

Well, contrary to ChatGPT’s best guess, neither violence nor alcohol were problematic, so what was common to the 2 stanzas that got bounced? Review, inspect, compare, aha! They both had the standalone word “dink” in them! So the platform’s hangup turned out to be excessive prudery. But a verse using “Dinkers” was OK. So I changed “dink away” to “dinkaway” and “dink-and-drop” to “dinkan drop” and reran those verses individually, and sure as shit, they passed muster.

= = = = = =

OK, time to deliver the entire wording (now up to V.5), minus the initial title and credits, to Suno. And the result was pretty satisfying, except that there were a couple of places where the singer would draw out the final word of a stanza (like “oh myyyyyyy”, “snacks and beeeeerrr”, “cooooollllerrr”, and “sslllooowwww”). Such stylings might be just fine for radio or a concert performance but they’re horrid if you’re trying to dance to it. For that you want a nice, steady, consistent, predictable rhythm. So I modified my style specs to read “German style polka, 2/4 time, oom-pah, accordion, cheerful, moderate pace, strict tempo, danceable, male and female vocalists” and sent it back again.

Next result was better in that regard, but I noticed that the normal pronunciation of “again” didn’t rhyme with “refrain” (despite the “ain” ending of both words), so I changed the spelling of the former to “agane” and also swapped out “teammate” for “partner” and “scorekeeper” (non-existent in pickleball) for “ref”. Then I prepared to feed it in once more (or “agane” if you prefer).

But as I got set to do so, I picked up on an aspect of the user interface that I hadn’t noticed before. It gave me the option to select “Vocal Gender” as either Male or Female, and I wanted both, so I tried to click on both, but it only let me pick one of them at a time. Fortunately, since I had requested dual vocalists in my custom specs, I didn’t need to do it here, and I was able  to cancel my choice by holding down the shift key and clicking on it again.

Encouraging news: It took Suno less time to create the two versions of the song than it took to play either (2½ and 3 minutes).

= = = = = =

So after all of that, here is V.5 of the lyrics, provided for your viewing pleasure.

🎶 The Pickle Barrel Polka 🎶

(in the cheerful beer-hall polka style of Frankie Yankovic and His Yanks doing “Beer Barrel Polka”)

Credited to: Yankee Frankovic and Half a Dozen Chubby Belgians

[Verse 1]

Come gather round the baseline, friends, the score is oh-oh-two,

We tap our paddles proudly — we know just what to do!

Keep your eye upon the ball, don’t pop up like a loon,

Unless you want your partner to eat it mighty soon!

[Chorus]

Roll out the pickle barrel, roll out the nets today,

Step up to the kitchen line and gently dinkaway!

We’ll chuckle and we’ll rally till sunset paints the sky,

And dream we’re champs like Anna Leigh and Ben, oh my!

[Verse 2]

At Garner Park in Madison the volleys never cease,

McGaw in Fitchburg echoes as the whack-whack-whacks increase,

At Taylor Park in Middleton, the players all agree,

This game’s more fun than fishing, twice as social as TV!

We practice third-shot drops with elegance and grace,

But sometimes someone slams it — oooh, try to miss my face!,

“Game to eleven, win by two!” the ref will cry,

No game will last forever, nor ending in a tie.

[Chorus]

Roll out the pickle barrel, let’s paddle side by side,

Stacking near the service box with partners full of pride,

No matter what your DUPeR — whether av-er-edge or super —

We toast the perfect rally with a dip inside the cooler!

[Verse 3]

Inside at Dill Dinkers when the winter winds blow cold,

Pickleball Kingdom’s crowded — each court is filled and bold,

Pickle Pro Courts are buzzing, Hitters hums with cheer,

CAPA friends all gather: “Who brought snacks and beer?”

We guard the kitchen fiercely — no volleying inside it!

But when circumstances make us step half an inch bestride it,

“Foot fault!” someone whispers, but no one makes a scene,

Because sportsmanship and laughter rule this paddle-happy green.

[Chorus]

Roll out the pickle barrel, sing a dinkan drop refrain,

We’ll rally till the lights go out then start it up agane,

From beginners to the dreamers who chase a champion’s fame,

We all just love the friendship and the joy of this fine game!

[Finale]

So raise your paddle high, friends, let the polka band play on,

We’ll chase that yellow wiffle till the daylight’s nearly gone,

And though we’re sore tomorrow and walking kind of slow,

We’ll be back again at dawn to shout — “Who wants to serve? Let’s go!”

= = = = = =

Since Suno was still throwing in arhythmic “stylings”, I hoped that Abigail’s subscription to OpenMusic.ai would be more accommodating, so I asked her to give it a spin. And that put the finishing touches on the song itself. It clocked in at 2:29.

OK, song generated, now to get it published. So I asked Pi.ai “Once a song-writing AI like Suno, Udio, or OpenMusic generates a song I like, how do I go about making it available on line to anyone who’d like to play it?” There was a wrinkle in the response:

You can share it online by downloading the audio file (if the platform allows) and uploading it to services like SoundCloud, YouTube, or Bandcamp. These platforms let you publish music for free and share links widely. Just make sure you understand the platform’s terms—some AI tools restrict commercial use or require attribution. For broad access, YouTube or SoundCloud are great choices.

Depending on the platform and how you’re using the song, ethically and often legally you should disclose if a song was AI-generated, especially if sharing publicly or commercially. Platforms like YouTube, Spotify, and SoundCloud are starting to require AI disclosure, and US regulators (like the FTC) recommend transparency to avoid misleading listeners. Some AI music tools (like Suno) also require attribution in their terms. When in doubt, credit the AI tool used (e.g., “This song was created using Suno.”) It’s honest, builds trust, and keeps you compliant.

Since I was familiar with YouTube, I tried that first, but my .mp3 file was audio-only and not a supported format, so Pi aimed me instead at 3 sites “made for music uploads [that] accept MP3s directly”:

 • SoundCloud is great for sharing widely and embedding links

 • Bandcamp works well if you want to offer downloads or even sell the track

 • Mixcloud is ideal for longer sets or curated playlists

All let you add AI disclosure in the description. Easy, free, and music-focused.

Right! So on to SoundCloud, where I created an account and said I wanted to upload a tune as an unpaid user. After pointing it at my file “Pickle Barrel Polka.mp3” it asked me for more info, and this is what I gave it:

 • Track title: Pickle Barrel Polka

 • Track link: https://soundcloud.com/richardsrussell/ (←pre-entered) picklebarrelpolka (←typed in)

 • Main Artist(s): Lyrics by ChatGPT from specs by Richard S. Russell; music and vocals by OpenMusic.ai

 • Genre: German polka (not one of their default options, so typed in)

 • Tags: #polka #2/4 time #accordion #male and female vocals #cheerful #upbeat #pickleball #fun #Wisconsin (tags separated by spaces)

 • Description: A jolly polka (purportedly by Yankee Frankovic and Half a Dozen Chubby Belgians) celebrating good times playing pickleball around Madison, Wisconsin, home to the Capital Area Pickleball Association (CAPA); the polka is the Wisconsin state dance. Generated by AI on 2026 Feb. 28.

Track Privacy: Public (other options were Private and Schedule)

Clicked “Upload” and ipso quicko there it was, ready to play:

https://soundcloud.com/richardsrussell/picklebarrelpolka

While listening to it for the first time, I noticed that it also provided me with the option to upload a picture of the account owner (me), so I did that, too. Then I bookmarked the track for future reference.

Now all I have to do is post the above narrative on my blogs and notify my fellow pickleballers of where to look at what I’ve done.