R to @itsanderz: Hi Andrew thank you so much 🙏

R to @cgonfire: Hey

RT by @itsanderz: What's better than one Nuke release? Two! So why now for our first-ever dual release? Let us explain 🧵

R to @Mrboofyy: This looks mad cool! What was the original video ?

R to @itsanderz: Heck yes, so sick!

R to @onlyfootnotes: Yeah, like I am from Jamaica and I didn't even know those stations existed and yes they are actual local stations and it's so MAD to me. I wonder how it all works.

R to @itsanderz: I know right??

R to @onlyfootnotes: This is so sick!

R to @itsanderz: @abacaj was looking to train/ fine-tune a truly open source text to music model on a dataset I curated. Is that something you could give me some direction with?

R to @abacaj: Oh, quite interesting... Will look into it! Should have specified text-to-music/ audio 😭

R to @itsanderz: I did actually try to FT this https://github.com/rhasspy/piper, more as an experiment though (not anything for deploying)

R to @abacaj: @abacaj have you tried training/ finetuning audio/ music models ?

R to @itsanderz: Yeah! I’ve tried it a bit on Huggingface. I havent’ installed it locally yet. I’m also familiar with the paper Meta released alongside with it and their examples are an impressive leap forward. I’ve been following it and similar tools closely. For example Suno AI, Stable Audio, JEN-1. I don’t think they’re any close to making one of the next top100 hit songs of the year. But as a music producer i think they can be very helpful for brainstorming and avoiding writer’s block. The most interesting more “professional” current use case for me would be in generating royalty free samples to make urban music instrumentals (specially rap/trap), since 5-10 seconds loop can sometimes be enough to be the backbone of a full length songs in genres where the main focus tends to be around lyrics over melody or music complexity. I know a lot of awesome rappers who struggle to find royalty free high quality instrumentals to write songs they can confidently publish on streaming services without the fear of copyright claims. Tools like these can be a game changer for those who can rap like eminem but can’t afford Dr. Dre haha (i mean can’t afford a decent producer) and allow them to reach a much wider audience. It also can be very helpful to show a producer a much better idea of the kind of sound you are looking for without the need of wasting their time (and your money) trying to understand what you mean. Just by learning basic prompting you could find exciting ideas on your own time and then simply learn how to make a basic loop out of it and record your vocals yourself, or giving it as a nice starting point for your producer to finish the song in a more straight forward way. I’ve been a musician, rapper and singer for a long time, but that difficulty was in fact one of the main reasons why I decided to learn how to be a producer. What’s your take on musicgen or the current state of these tools? PS: I’ve been using more narrow music AI tools for some time, such as Ozone’s plugin for automatic mix and mastering once you have all your tracks already recorded. Also some promising EQ plugins. Overall I think this is similar to programming with ChatGPT or GitHub Copilot. They can lower the barrier of entry A LOT. But the more you know the field, the more impressive end result you’ll get (or time you’ll save). It’s certainly not replacement for programmers, musicians or producers but they can quickly become an unfair advantage if you know how to integrate them into your workflow.

>@Dis_Trackted Hey, have you tried the musicgen model?

R to @levelsio: @levelsio do you have to connect replicate to Lambda? Also, do you train on replicate as well?

R to @carltongstew: Where should applications be sent ?

R to @abacaj: @abacaj have you done this with audio as well ?

R to @itsanderz: The differences picked don’t match the picture though?!

R to @AndreTI: I love the mouse’s tail. These things might be mostly wrong linguistically, but if you begin to reinterpret what some of these words mean, it’s an added clue. Maybe models will become more precise. But seeing the cat with a lollipop being interpreted that way is actually adorable

R to @itsanderz: Are you saying we can find Wally now???