I asked #Udio to make a #DepecheMode song and here's what happened!

Up till now, Suno.com has been the big thing in AI music generation, but now there's a new kid on the block called Udio.com. I gave it a whirl and it was interesting to see how it behaved when you put in a band name as a prompt. Suno won't let you do that, but Udio handles it in a smarter way by replacing the band name with suitable prompts: https://www.andrea.net/temp/Udio_How_it_handles_band_names.jpg

So I tried "Depeche Mode" and here's what happened! * blown away *

Since I've had a lot of questions and some people even doubting as to whether this was made with Udio, here is some more detailed info:

The original prompt I used was "Depeche Mode". I did not add any lyrics although Udio allows you to do that. Instead I let it just generate it's own lyrics (which is why they are so bad). In a way though, I think it's kind of cool because it makes it a purer experiment of what the software can do.

Whenever you generate music in Udio, it will spit out two sample clips. When I generated this song, it's "twin" was this: https://www.andrea.net/ai_music/Sync%20in%20the%20Veneer.mp3
Although it has a lot of potential, it also has a lot of problems and does not sound as good or as much like Gahan, so in that moment I went for the one you hear in the video.

I was so excited by "the better twin" that I immediately got to work adding extensions to make it a complete song. It took hours and hours because 99% of the generations did not work well together with the initial clip. If I remember correctly, most of the extension clips did however sound like Gahan - but at varying degrees.

When generating the extensions, I never touched the prompts and the initial "Depeche Mode" was still in the prompt field. Sometimes Udio replaced "Depeche Mode" with an identical prompt, but sometimes there was variation. The final song ended up consisting of 6 parts that you can see me browse through here: https://www.youtube.com/watch?v=I8dDs7pihQc

Here are the prompts Udio replaced "Depeche Mode" with for the 6 parts:
1. Male vocalist, Electronic, Synthpop, Rock, Alternative rock, Dark, Rhythmic, Melodic, Atmospheric, Nocturnal, Alternative dance, Mechanical, Passionate, Cold, Longing, Sombre
2. Male vocalist, Electronic, Synthpop, Rock, Alternative rock, Dark, Rhythmic, Melodic, Atmospheric, Nocturnal, Alternative dance, Mechanical, Passionate, Cold, Longing, Sombre
3. Male vocalist, Electronic, Synthpop, Rock, Alternative rock, Dark, Rhythmic, Melodic, Atmospheric, Nocturnal, Alternative dance, Mechanical, Passionate, Cold, Longing, Sombre
4. Instrumental, Instrumental, Electronic, Synthpop, Rock, Alternative rock, Dark, Rhythmic, Melodic, Atmospheric, Nocturnal, Alternative dance, Mechanical, Passionate, Cold, Longing, Sombre
5. Instrumental, Instrumental, Instrumental, Electronic, Synthpop, Rock, Alternative rock, Dark, Rhythmic, Melodic, Atmospheric, Nocturnal, Alternative dance, Mechanical, Passionate, Cold, Longing, Sombre
6. Instrumental, Instrumental, Instrumental, Instrumental, Electronic, Synthpop, Rock, Alternative rock, Dark, Rhythmic, Melodic, Atmospheric, Nocturnal, Alternative dance

Once the end result was satisfactory and I had finished this video, I shared it on the Udio discord and I know that at least one of the devs was around at the time.

The next day, I tried making more songs by writing in "Depeche Mode" and letting Udio do it's replacement thing. But no matter how many times I tried (and have kept on trying since), I was never able to get these kinds of results. Some vocals sounded like Peter Murphy, some with unknown gothy voices and only once in a while would I get a song where the vocals sounded "a little bit" like Gahan.

It feels to me like Gahan's voice was one of the many that they used to train Udio on and that it was coded in such a way that it would randomly mix the vocals from different bands with varying amounts. And that this song just happened by some fluke chance to have a very high amount of Gahan. The "twin" clip I linked above maybe has 80% Gahan. One of the songs I generated today had maybe 50% Gahan. I mean, I have no idea how they've got it all set up or how this works, but that's just what it feels like.

So I mentioned the dev being in Discord at the time that I shared this video. Is it possible that they saw the commotion about how much this sounded like DM (when people were commenting on it) and alerted the team that some tweaking needs to be done? I have to wonder this since it's been impossible to get the same results ever since.

PS: The images are generated with #StableDiffusion and are meant to look like #DaveGahan but this is merely the interpretation of one of the many models you can find on #civitai so the likeness is not perfect. I couldn't be bothered training his likeness or anything like that.

#aimusic #aimusicgenerator #suno #udio #udiomusic #aimusicevolution