Results from running randomly-generated neural networks (all weights in neuron connections start random and then slowly drift when generating). Although sample rates used are 44.1 and 48 kHz, sounds pretty lo-fi. The reason could be input compression needed for network to actually work. Each sound channel is an output from two separate neurons in the network. Each sample in this pack is generated by a separate network, as they wasn’t saved anywhere after they produce a thing. Global parameters (output compression, neuron count, drift rate etc.) aren’t the same from sample to sample, too.
Type
Flac (.flac)
Duration
0:10.000
File size
1.3 MB
Sample rate
41100.0 Hz
Bit depth
16 bit
Channels
Stereo
Comments
7 years, 3 months ago
P. S. The algorithm had minute errors, here’s an updated description: https://pastebin.com/raw/yXrxMqT2
7 years, 3 months ago
@NightDen Yeah, definitely something. I think both these ones and sounds from the Matrix are somewhat dial-upish. Well, 90s!
7 years, 3 months ago
Sounds like the matrix :)
8 years, 6 months ago
Here’s the algorithm, if someone’s interested.
8 years, 6 months ago
Glad to hear that! (I wonder, what did they do for that sound, then?) I mean, sounds from this pack was quite déjà vu to me, too, from the start—something 90s-ish—and people back then should have means far more simple to create that. Neural nets is definitely an overkill. (OTOH that day I’d tried to use something like IIR filter, also with random coefficients and drift, and very subtle noise for input, but it sounded far more uninterestingly unpleasant that nets; it seems there was just not enough hidden state?)
- 1
- 2
