4 ms·

So how long before we can “Hey music-stable-diffusion-service,” play me “relaxing baroque violin and flute music in the style of game of thrones” ?

echelon 4y ago
My start-up is working on this! We'll be launching a web and downloadable version soon. Eg. https://www.youtube.com/watch?v=r_0JjYUe5jo https://www.youtube.com/watch?v=r_0JjYUe5jo --> https://vocaroo.com/1hgjjnVNqWjk https://vocaroo.com/1hgjjnVNqWjk We're also working on film generation.
- mjr00 4y ago
  Not to insult your startup or your work (as this is pretty cool), but this is a good illustration of why that last 10% is so important. While the AI-generated song is 90% similar at a surface level, there's also no way anyone would listen to that track on Spotify if it came up on an Eminem playlist.
- o_____________o 4y ago
  I don't understand what I'm seeing here, in ingested the Eminem track and recreated it using a different voice?
anonymouse008 4y ago
I think we all can view a video of 'nails on a chalkboard' before we hear audio of nails on a chalkboard. For some reason, unacceptable uncanny sounds is a much wider valley than unacceptable videos/pictures. The hand holding is uncanny in the family video, but I'm fine watching it for a second - it doesn't cause pain the way that same error would in music.
- colordrops 4y ago
  I recall AI generated Beethoven years ago that sounded great. What happened to that research?
ragazzina 4y ago
I've found this is my biggest issue with music streaming services. Often I just want another song similar to the one I am listening to, but no "generate radio from this track" comes close.
mjr00 4y ago
A long time. Much like self-driving cars, AI can take you 90% of the way there, but the last 10% is the difference between music that sort-of-seems-competent, and music that people will actually listen to and enjoy.
- consumer451 4y ago
  Is it really the same though? A driving system which can only create a good outcome 90% of the time is not really useful as irl human safety is a factor. However, a creative system curated by a human could end up creating useful outputs, could it not?
  - zild3d 4y ago
    100%, poor music output can be responded to with "next". poor self driving car output can be a collision
  - mjr00 4y ago
    Obviously different safety outcomes, but ultimately it's the same idea: unless you can hit that last 10%, it's essentially useless. You could pump out AI songs that are about as good as the 10,000 songs that get uploaded to Soundcloud every day, but nobody listens to those already. It's really only the best <1% of music that people listen to. Something that can be further refined by humans is more interesting. There's people looking into AI-based sample generation which is a lot more promising than full song generation, IMO.
    - consumer451 4y ago
      > Something that can be further refined by humans is more interesting. Exactly, that is what I was trying to say. The way I look at it is that most people who have Ableton installed cannot create an amazing song. Now let's say they are able to prompt a Stable Diffusion Audio system with a prompt like kanye type beat with flute melody in the key of E. The system might output 90% hot garbage, but it's easy to skip that within seconds of hearing it. So they clip and loop the good part, add whatever personal skills they do have, and upload that. And wow, I just found out that OpenAI's Jukebox[0] was creating this stuff two years ago. This seems like the lowest hanging fruit to me, compared to visuals. Also could be extremely lucrative. I wonder if we are already listen to ML generated music and it's just not advertised? [0] https://openai.com/blog/jukebox/ https://openai.com/blog/jukebox/ related post: https://news.ycombinator.com/item?id=23032243 https://news.ycombinator.com/item?id=23032243
anigbrowl 4y ago
That's so last week: https://mubert.com/ https://mubert.com/ (not affiliated)

echelon 4y ago

My start-up is working on this! We'll be launching a web and downloadable version soon. Eg. https://www.youtube.com/watch?v=r_0JjYUe5jo https://www.youtube.com/watch?v=r_0JjYUe5jo --> https://vocaroo.com/1hgjjnVNqWjk https://vocaroo.com/1hgjjnVNqWjk We're also working on film generation.

mjr00 4y ago
Not to insult your startup or your work (as this is pretty cool), but this is a good illustration of why that last 10% is so important. While the AI-generated song is 90% similar at a surface level, there's also no way anyone would listen to that track on Spotify if it came up on an Eminem playlist.
o_____________o 4y ago
I don't understand what I'm seeing here, in ingested the Eminem track and recreated it using a different voice?

anonymouse008 4y ago

I think we all can view a video of 'nails on a chalkboard' before we hear audio of nails on a chalkboard. For some reason, unacceptable uncanny sounds is a much wider valley than unacceptable videos/pictures. The hand holding is uncanny in the family video, but I'm fine watching it for a second - it doesn't cause pain the way that same error would in music.

colordrops 4y ago
I recall AI generated Beethoven years ago that sounded great. What happened to that research?

ragazzina 4y ago

I've found this is my biggest issue with music streaming services. Often I just want another song similar to the one I am listening to, but no "generate radio from this track" comes close.

mjr00 4y ago

A long time. Much like self-driving cars, AI can take you 90% of the way there, but the last 10% is the difference between music that sort-of-seems-competent, and music that people will actually listen to and enjoy.

consumer451 4y ago
Is it really the same though? A driving system which can only create a good outcome 90% of the time is not really useful as irl human safety is a factor. However, a creative system curated by a human could end up creating useful outputs, could it not?
- zild3d 4y ago
  100%, poor music output can be responded to with "next". poor self driving car output can be a collision
- mjr00 4y ago
  Obviously different safety outcomes, but ultimately it's the same idea: unless you can hit that last 10%, it's essentially useless. You could pump out AI songs that are about as good as the 10,000 songs that get uploaded to Soundcloud every day, but nobody listens to those already. It's really only the best <1% of music that people listen to. Something that can be further refined by humans is more interesting. There's people looking into AI-based sample generation which is a lot more promising than full song generation, IMO.
  - consumer451 4y ago
    > Something that can be further refined by humans is more interesting. Exactly, that is what I was trying to say. The way I look at it is that most people who have Ableton installed cannot create an amazing song. Now let's say they are able to prompt a Stable Diffusion Audio system with a prompt like kanye type beat with flute melody in the key of E. The system might output 90% hot garbage, but it's easy to skip that within seconds of hearing it. So they clip and loop the good part, add whatever personal skills they do have, and upload that. And wow, I just found out that OpenAI's Jukebox[0] was creating this stuff two years ago. This seems like the lowest hanging fruit to me, compared to visuals. Also could be extremely lucrative. I wonder if we are already listen to ML generated music and it's just not advertised? [0] https://openai.com/blog/jukebox/ https://openai.com/blog/jukebox/ related post: https://news.ycombinator.com/item?id=23032243 https://news.ycombinator.com/item?id=23032243

anigbrowl 4y ago

That's so last week: https://mubert.com/ https://mubert.com/ (not affiliated)