Conversation

Notices

1. Embed this notice
   Lorewegian 📚 (lore@berserker.town)'s status on Monday, 22-May-2023 19:07:13 JST Lorewegian 📚

   in reply to

   • iced depresso
   • ringo

   @icedquinn @ringo i saw what you were going for, but sometimes, i wonder what would happen if you just had the computing power to run a neural network in real time to generate the output directly. might produce some fairly wild sounds.

   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 22-May-2023 19:07:14 JST iced depresso

     in reply to

     • ringo
     @ringo @lore i'm just talking about auto-tuning an antialiasing filter so you can get as close to heavy oversampling as you can without having to pay for it
     
     so most of these objections are just not understanding what i said
   • Embed this notice
     ringo (ringo@talk-here.com)'s status on Monday, 22-May-2023 19:07:16 JST ringo

     in reply to

     • iced depresso
     @icedquinn @lore purposeful technique over "pasta sticka to the wall now?"
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 22-May-2023 19:07:17 JST iced depresso

     in reply to

     • ringo
     @ringo @lore i think people are not understanding you are just using the numerical optimizer to tune the constants for the much cheaper filter. you aren't putting an AI on chip
   • Embed this notice
     ringo (ringo@talk-here.com)'s status on Monday, 22-May-2023 19:07:19 JST ringo

     in reply to

     • iced depresso
     @icedquinn @lore
     
     interesting. just my 2 cents but keep ai out of the hardware design.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 22-May-2023 19:07:20 JST iced depresso

     in reply to

     • ringo
     @lore @ringo well something fun is anything that has a derivative in calculus can be tuned with the same stuff we make AIs with.
     
     so like i said you can have a script generate hours of audio at random settings from a synth and like.. have it compare the aliased to 16X oversampled waveforms, and crunch a biquad filter to get as close to the output as desired. or you can add/remove terms to the biquad filter and see if thats better.
   • Embed this notice
     Lorewegian 📚 (lore@berserker.town)'s status on Monday, 22-May-2023 19:07:21 JST Lorewegian 📚

     in reply to

     • iced depresso
     • ringo

     @ringo @icedquinn IIRs being recursive is one of the reasons they're difficult to parallelise. but they are, musically speaking, pretty nice sounding.

   • Embed this notice
     Lorewegian 📚 (lore@berserker.town)'s status on Monday, 22-May-2023 19:07:22 JST Lorewegian 📚

     in reply to

     • iced depresso
     • ringo

     @ringo @icedquinn must be some kind of approximation since IIRs are recursive.

   • Embed this notice
     ringo (ringo@talk-here.com)'s status on Monday, 22-May-2023 19:07:24 JST ringo

     in reply to

     • iced depresso
     @icedquinn @lore
     
     interesting.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 22-May-2023 19:07:25 JST iced depresso

     in reply to

     • ringo
     @ringo @lore there was an article someone posted on an audio forums about "mini-FIRs" but i don't think i ever saved it properly.
     
     was very math heavy though. but was something to do with reconstructing analogue filters instead of reverbs with impulse responses.
   • Embed this notice
     ringo (ringo@talk-here.com)'s status on Monday, 22-May-2023 19:07:27 JST ringo

     in reply to

     • iced depresso
     @lore @icedquinn are you familiar with this?
     
     https://www.korg.com/us/products/synthesizers/modwave/index.php
   • Embed this notice
     Lorewegian 📚 (lore@berserker.town)'s status on Monday, 22-May-2023 19:07:29 JST Lorewegian 📚

     in reply to

     • iced depresso
     • ringo

     @icedquinn @ringo you end up with some pretty funky compromises. i needed to have a sine waveform for the synth but it sounded digital if i made the LUT too small.

     then i discovered that if you add some dithering noise to the LUT index, you get hiss instead, which is much better. and you can easily filter that away, leaving a nice sine wave.

   • Embed this notice
     Lorewegian 📚 (lore@berserker.town)'s status on Monday, 22-May-2023 19:07:37 JST Lorewegian 📚

     in reply to

     • iced depresso
     • ringo

     @icedquinn @ringo LUTs are absolutely necessary for adequate performance, but you risk them getting too big if you require some precision, and then you exhaust the cache memory in the CPU.

   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 22-May-2023 19:07:38 JST iced depresso

     in reply to

     • ringo
     @lore @ringo somehow feel like you could cheat all this with some kind of lookup table filter with optimized constants
   • Embed this notice
     Lorewegian 📚 (lore@berserker.town)'s status on Monday, 22-May-2023 19:07:49 JST Lorewegian 📚

     • iced depresso
     • ringo

     @icedquinn @ringo arguably, that's 12 multiplications when you oversample it 4X. but in theory you could get away with 2X if you had a very steep anti-aliasing filter. unfortunately, such a steep filter is going to cost you *way* more multiplications...