To the new playlist

My new AI music project is called “Shadow Light goes Irish.” For this project, I’m producing a playlist of about 10 tracks featuring songs in the Zurich German dialect in an Irish folk and rock style. It sounds great, even if the dialect doesn’t always come across perfectly. Even among people in a big city like Zurich, the Swiss dialect is already somewhat worn down and interspersed with a mix of other dialects. The dialect of the artificial singers is also blended with other dialects. You can’t help but notice this when you play these songs—a minor drawback you have to accept with songs in dialect.

What Works—and What Doesn't

On the positive side, however, the Irish musical style opens up a new musical dimension for dialect songs, one that fits very well with the raw and rebellious character of Irish folk music. Nevertheless, there is an insurmountable hurdle here as well: You can’t simply cover Irish music by well-known groups like the Dubliners or sing along to translated versions in the typical Dubliners style so that the dialect song sounds exactly like the original.

Copyright law alone sets strict limits here. A great many songs by Irish music groups are protected, as is the music that accompanies them. So you can’t just take a folk song and translate it into German in a way that makes it sound like the original song. Music by well-known groups like the Dubliners is so distinctive that you can always recognize it as such.

But what you can try are folk songs—songs that are part of the traditional repertoire and don't have specific authors. There's an example of that in my new setlist, too.

So I tried to produce a well-known Irish song in Swiss German: “The Black Velvet Band.” I used Suno to do it. Almost everyone knows the melody and the rougher sound of the Dubliners’ version. But the result of my new song is a world away from the original Dubliners song. Hardly any of the listeners believed me when I said my intention was to re-record this song. Everyone commented that while the lyrics were similar and it sounded “Irish,” it was a completely new song featuring the “samtigä Band.” Hardly anyone made the connection to the original “The Black Velvet Band.”

But even if the lyrics are in the public domain, that doesn’t mean you can use the Dubliners’ voices for free. Even if you don’t directly sample the Dubliners (i.e., use audio snippets from the original in your own song), you’re still infringing on what’s known as the right to one’s own voice—for example, if an AI has been trained on Ronnie Drew’s rough, pub-style intonation and then uses it in a recording. A voice is considered a distinctive personal characteristic. There are historical precedents, such as the case of Bette Midler, where a singer was sued for imitating her in a deceptively realistic manner.

Major record labels like Sony and Warner are currently filing massive copyright lawsuits against AI apps like Suno. They accuse these apps of illegally training their AI models on millions of protected original recordings. It’s safe to assume that well-known music groups like the Dubliners are certainly among them. If the sound is too close to the original, competition and trademark laws also come into play. As a listener, you might think: “Great. That sounds just like a long-lost song by the Dubliners that’s now been released in Swiss German.” And you might assume that the original band authorized and supported this project. But that’s certainly not the case with AI-generated music.

Given this somewhat frustrating situation, I explained my problem to the AI bot Gemini once again in no uncertain terms. This was my request:

“Everyone thought it was a similar text, but a completely new song. No one associated it with the Irish song ‘The Black Velvet Band’ anymore. Isn’t there a way to arrange the song with Suno so that it retains the original melody and stays closer to the original?”

Here is the detailed and somewhat “technical” explanation, which, however, accurately describes the dilemma:

"This is a well-known and extremely frustrating problem when working with Suno. Unfortunately, the short, honest answer is: No, you can't (yet) force Suno to sing along to a specific, existing melody exactly note for note."

It's completely understandable that your audience didn't recognize the song. Suno doesn't work like a human musician who is given sheet music or a familiar song.

Here's why that happened, and a few strategies for how you can at least get much closer to the original "Black Velvet Band" in the future:

Why Suno "ignores" the original

• Pure text-to-music model: Suno generates music based on statistical probabilities derived from text prompts. When you translate the text into Swiss German, the number of syllables, the stress patterns, and the rhythm of the language change completely. Suno then composes a new song that perfectly matches the new flow of the Swiss German text.
• No "note entry": With Suno, you can’t upload MIDI files or sheet music to dictate the melody.
• AI Copyright Filters: “Although the traditional melody of ‘The Black Velvet Band’ is in the public domain (not subject to copyright), Suno’s AI models are trained to avoid exactly copying copyrighted material (such as the specific recordings by The Dubliners). Sometimes the system even proactively blocks exact melodies.”

Gemini also offers a few tips on how to get a little closer to the original. However, these can quickly become very time-consuming to implement. Only a professional producer—who could spend days working on such a recording—has that much time. And the result is still somewhat uncertain as long as voice privacy protections remain in place.

Gemini demonstrates how one could technically achieve a similar result by suggesting, for example, that you use the audio upload feature in this case:

• Suno offers an "Upload Audio" feature. You can upload a short clip (up to 60 seconds) from an existing audio file.
• The trick: Look for a royalty-free, purely instrumental version of “The Black Velvet Band” (or, if possible, play the distinctive melody yourself briefly on the guitar or keyboard).
• Upload this clip and select “Extend.” Paste your Swiss German text there. Suno will then try to apply the style, rhythm, and often the melody of the uploaded snippet to the rest of the song.

Regarding these and a few other techniques that were described, I shared the following conclusion with Gemini: “Even with these very elaborate methods, the song still hardly sounds like ‘The Black Velvet Band’ as we know it. I suppose I’d have to describe the new song more like this: ‘Traditional: based on the lyrics of “The Black Velvet Band.” Reinterpreted by Shadow Light Project.’”

The immediate reply:

“You’ve really hit the nail on the head. Your disappointment is completely justified, and the label you’ve suggested is spot on: That’s exactly the honest and artistically accurate description of what happened here.”

Your listeners are absolutely right: Anyone who has the classic, driving melody of the Dubliners in their head will search for it in vain in the AI version. Even with audio uploads and precise rhythm specifications, Suno behaves like an extremely idiosyncratic jazz musician—the AI might pick up on the chords or a vague stylistic direction, but then immediately breaks out into its own, completely unpredictable improvisation.”

After giving all this some thought, I’ve decided to stick with my original concept. For my “Playlist Irish,” I’m including ten songs in an Irish folk-rock style. These are adaptations of well-known Irish songs, but without trying to replicate the original sound. The lyrics of these songs are often so expressive that they remain moving and appealing even when given a new arrangement. In doing so, I try to modernize them slightly within the framework of their existing lyrics and, for example, tailor them to new circumstances.

So the Swiss German protagonist, Koni Koller—like many of his peers—sets off on a trip around the world after graduating from high school and, in New York, is so thoroughly tricked by the fateful girl with the black ribbon that he ends up in a U.S. maximum-security prison.

This project also features new Swiss German songs, which, in my opinion, benefit greatly from the Irish instrumentation. One example is the “Ossola Song,” which draws inspiration from the rebellious songs of the 1970s and captures in a single song the support for the Italian freedom movement against the German Nazi occupation.

And the conclusion

Overall, songs like these are more about building bridges between cultures and mutual enrichment than about trying to promote an understanding of Irish music through dialect-style cover versions. Gemini described it beautifully: It’s not a cover playlist, but rather an atmospheric collaboration between Irish folklore and modern AI culture, presented in the Zurich German dialect.

That could be the conclusion of this blog post. But things are moving so fast right now that what’s said here might no longer be true in just a few weeks. So, looking ahead, I’d like to share what Gemini told me about this new development: Google’s “Lyria” music model is taking a completely new approach:

Lyria (Google DeepMind's music model, which is being tested in YouTube Shorts Dream Track, among other places) and similar next-generation AIs take a fundamentally different approach than Suno:

• Voice and Melody Transfer: While Suno always composes from scratch, models like Lyria are designed to analyze existing audio characteristics.
• How this would solve your problem: With models of this generation, it will be possible to tell the AI: “Use the exact melody line and arrangement from recording X, but replace the voice with a singer from Zurich and adapt the vocal rhythm to the new lyrics.”

Lyria and similar advanced music AIs are often still in closed testing phases or are only accessible to select artists and platforms (such as YouTube tools) due to complex licensing and copyright negotiations with the major music labels. But this technology is exactly the bridge you’re looking for: it transforms existing material rather than reinventing it from scratch.