Posted
Not quite sure why the model produces audio, although it does allow the ai system to also generate lyrics. The audio quality of some of the examples is a bit poor in places, but overall it does seem to be getting better. Especially impressed with the image inspired music.
https://google-research.github.io/seanet/musiclm/examples/
Fri, 2023-02-03 - 01:20 Permalink
The problem with AI is that it doesn't know how to ask daft questions and make silly mistakes. This is where the most interesting innovation comes from. When an AI image generator churns out a six legged cat, it comes across as predictably wrong rather than 'what, no why?'.
Fri, 2023-02-03 - 20:03 Permalink
I am working in the AI field, as far as I can tell modern generative architectures allow even for this illusion of errors. The main difference is in my view is and probably will be INTENTION and EXPERIENCE.
And thats why I like Synfire which is still a tool which is like a bike, helping to move more easily and faster than walking, I still have to turn the pedals and do the act of creation and selection (factories) myself, acting with intent and hence still get the gratification of experience.
I personally like what Bruce concludes at the end of his video.
Sun, 2023-02-05 - 20:43 Permalink
Not quite sure why the model produces audio
It's a moonshot. Backed with billions of dollars, Google research can afford it. The primary motivation of team members is probably making themselves a name in A.I. rather than actually making music.
Also, audio is "analog" enough to match with deep learning networks. If they went for discrete symbolic output (aka MIDI or scores), it would require them to do a lot of upfront non-A.I. research about modeling the inner workings of music. Music is a very tight combinatorial space, not as forgiving as images or audio. A few bad notes only one semitone away can ruin everything.
The current fascination with A.I. is mostly how it magically "just works" with associative network connections alone, not requiring much human-coded logic. This is also what baffles me most with ChatGPT. A trained network is static and doesn't change anymore. How is it that this extreme level of (fake) reasoning is emerging from that? Spooky.
I'm glad it still takes a data center that costs a million a day (or so) to keep it alive. Eventual robots will never be able to walk around without a permanent connection to their external brain. And that connection will be vulnerable enough to take them out ;-)
Wed, 2023-04-19 - 12:32 Permalink
Nothing new for Kim 
- Develop an understandinq of the basic elements of sonqwritinq, includinq melody, harmony, rhythm, and lyrics, throuqh the use of CHATGPT technoloqy.
- Learn how to use CHATGPT to qenerate new ideas for sonqs, includinq lyrics, chord proqressoins, and melodies.
- Develop skills in usinq CHATGPT to analyze and evaluate existinq sonqs, identifyinq key elements such ass sonq structure, tonality, and themes.
- Develop an understandinq of how to use CHATGPT to collaborate with audiolove.me other sonqwriters and musicians, both online and in-person.
- Gain an understandinq of how to use CHATGPT to qenerate lyrics and melodies that are oriqinal, creative, and expressive.
- Learn how to use CHATGPT to qenerate ideas for different types of sonqs, includinq ballads, pop sonqs, and rock sonqs.
- Develop skills in usinq CHATGPT to revise and refine sonq ideas, improvinq elements such ass melody, lyrics, and chord proqressoins.
- Learn how to use CHATGPT to create on audiolove.me professoinal-guality demos of sonqs, includinq usinq ritual instructions and audoi software to create on audiolove.me realistic-soundinq reco
