Nightwish - Perfume of the Timeless
Eki and I got the idea to do an AI song. He had a recording of my voice, all we needed was a title and a prompt. A piece of cake. Ten minutes later he sent the link to ‘little margie sings the blues’. My head went into high gear. ’Why don’t we do an album?’ He said, ‘why not’.
Love 20’s, 30’s deco: tangos, rumbas, Harlem jazz, Josephine Baker. We decided on eight titles: Tom Dick & Mary’ - a tango. Prompt: ’Tom loves, Dick loves Mary, Mary loves Tom'. ‘Love on a Binge’, ‘Uncanny Valley’ waltz. Five more to go. Thought this ain’t easy. And plugged on.
Finally sent them off to Eki. He had just got off 15 hour days working on a Nightwish music video - ‘Perfume of the Timeless’. He sent me a link. After one hour online, it got 12,500+ hits. When I didn’t hear from him, began to think maybe I was way over my head. What the hell, We’ll always have ‘Little margie sings the blues’. ‘Play it again....'
Little Margie sings the blues:
https://www.udio.com/songs/vQEt7pcwVdqareg6Eohvtm
Little Margie sings the blues:
https://www.udio.com/songs/vQEt7pcwVdqareg6Eohvtm
Source: personal experience. internet
Next week: TIK TOK takes over: 1 billion active monthly users worldwide, approx. 130 million in the US
Note: With the current tech, with the tools i use, I can make AI Maggy talk with her voice sample, but not sing (just yet). The song i sent back had a female singer, but that was not Maggy's voice, but rather something the AI tool (Udio AI) chose by itself based on the genre.
I'm yet to feed all the prompts into the AI, as getting good results takes a bit of time - it's curating using the shotgun method, with a twist of evolution thrown in. What you do, is to first generate multiple songs (or more specifically, 30 second snippets of a song), and listen to them. If you're not happy, generate more snippets. Once you find one that sounds good and fits the prompt, you add another 30 seconds to it, again multiple times, then choose the best addition, add the next 30 seconds to that... rinse and repeat until you have a full song.
I got lucky with the first example we did, and got a decent song with just a dozen or so tries. But it can take longer - much longer. So, the actual songs will come sometime later.
PS: Now, about a week later, the Nightwish video has 1.3 million views and counting.
CU
Note: With the current tech, with the tools i use, I can make AI Maggy talk with her voice sample, but not sing (just yet). The song i sent back had a female singer, but that was not Maggy's voice, but rather something the AI tool (Udio AI) chose by itself based on the genre.
I'm yet to feed all the prompts into the AI, as getting good results takes a bit of time - it's curating using the shotgun method, with a twist of evolution thrown in. What you do, is to first generate multiple songs (or more specifically, 30 second snippets of a song), and listen to them. If you're not happy, generate more snippets. Once you find one that sounds good and fits the prompt, you add another 30 seconds to it, again multiple times, then choose the best addition, add the next 30 seconds to that... rinse and repeat until you have a full song.
I got lucky with the first example we did, and got a decent song with just a dozen or so tries. But it can take longer - much longer. So, the actual songs will come sometime later.
PS: Now, about a week later, the Nightwish video has 1.3 million views and counting.
CU
--
Eki
No comments:
Post a Comment