@jasonw22 https://github.com/chrisdonahue/wavegan - this looks promising for neural network based audio generation, but unfortunately my integrated intel graphics card doesn’t allow me to do anything with this (but I could maybe try spinning some spell.run instance with it);d There is also a JS generator for trained network so the final version should be easy to embed on website and I could help with some nice generative visualizations in the site background.
Thanks @karol ! I’ll take a look, this looks promising.
Earlier explorations ( Postcards from slow places - a lines compilation of long-form ambient music ) indicated that perhaps RNN would work better than GAN on music-as-audio (as opposed to music-as-MIDI) but this looks like a different approach to GAN, potentially.
AWS makes available some pretty hot-stuff machines for this sort of thing if we end up needing more horsepower. I’m willing to invest a few pennies into learning more about this approach, if necessary…
So far the hardest thing is just getting a machine working that has all the necessary libraries installed correctly… oddly, I have a feeling the actual neural network training and production of new music will be pretty easy once that bit is done.
I have this fantasy of continuously generating new works and hosting them on a liquidsoap streaming radio station, as @dailybells suggested.
Instead of a sample-by-sample approach, you could try to simplify a little bit by training on feature analysis data of larger bits of audio – wavesets of some size where the slices are always at zero crossings might be useful and be able to retain stuff like pitch information. Then you could do the reconstruction by concatenating wavesets instead of trying to resynthesize sample-by-sample… I’m working on something like this for another project at the moment. I’m more interested in using ML to label the wavesets for later resynthesis (as in, I’m not at the moment looking to train a system to reconstruct a signal, I’d like to get it to organize a collection of little bits of audio based on a large number of features extracted from them so I can play with the dataset in a more directed way later but maybe this approach could be useful for your project too?)
It very well may be! I’d love to learn more about what you’re up to, it sounds great.
I’ve been working with the analysis data segmented on onsets for a couple months. Piles of renders but not much worth sharing yet I don’t think… I just started switching over to using waveset segmentation instead of onset detection (though I’m keeping both approaches in mind as I think a mixture of approaches might be useful for what I’m working on…) but I wrote about my initial steps with the previous approach a bit here: http://blog.southseas.fun/synthesizing-environments.html
This is beautiful.
i am truly astounded. 34 hours! thank you everyone, i’m very much looking forward to a (long) listen.
My piece, “I Don’t Need Anything from Here” is made up of recordings created over the last year, many of which I had left stranded in forgotten corners of my hard drive until 6 weeks ago…a combination of old sketches, nearly finished compositions and field recordings. About 20-25 of these were culled in a sprint lasting about a week, after which I used them as building blocks to create the two halves of the finished track. The main tool I used in the final recording stretch was my ER-301, though I can’t recall too many patch specifics; just a lot of feedback loopers, noise and sine waves for the first half especially. Final assembly was done in Ableton over a couple of weeks, and - to my complete and utter shock - the process didn’t have me running for the hills.
I’m really happy with how the piece turned out – feels like I’ve crafted a sense of narrative that sustains itself reasonably well. Here’s a little teaser, combined with ancient images of mine from a previous life as a filmmaker…they seemed to fit well.
I have made an audio “index” for our postcard compilation. I split every track into two minute sections, took the two minute passage that contained the very centre point of its original track, applied a thirty second fade in and fade out, then stitched all 55 passages together in album order (so that as one song begins to fade out, then next song begins to fade in). The result is an 83 minute overview of the postcard compilation. It flows nicely, and shows that we really covered some diverse terrain.
On the off-chance that anyone would like a copy, I’ve put the FLAC file in Dropbox:
If anyone downloads it, please send me a note so that I know.
At the same time, I realise that we’ve all got a big playlist already waiting for us … so I understand if there are no takers!!
I am deeply grateful for the rich and seemingly endlessly surprising resource that is this community!
Notes on my track: Self Organising Forms - It started with a generative Max patch, recording some note data over to Ableton. I then did a bit of tinkering there - adjusting tempo, scaling etc. Lots of LFOs. Several takes with different instruments - Collision and Operator. An Ablteon device I built with chucker~ (the backwardsy stuff). A lot of shifting audio against each other. Really trying to make computer stuff not too perfect these days. Then a bunch of processing through the modular. This takes so long with a 90min track ie: recording an effects pass. I probably did ten or so full passes on separate instruments for this.
I like how it turned out, though I still find trying to do anything minimal very hard. I think I failed here in that respect! but happy with it nonetheless.
Notes on my track: Imported an Ambient modular performance from Lucid Grain (my collaboration with Martha Bahr) into the Ambient V0.3 software and let it run for 3 hours. Edited the most interesting parts.
Some notes on my track ‘For N.M.’
I took the patch apart shortly after recording, so these are rough notes. The only sound source was the 4ms SMR through Erbeverb, then straight into my interface. The SMR was getting pings on both inputs from Bastl Little Nerd in Euclidean mode, clocked by a really slow LFO (either Batumi or uLFO, don’t quite remember). Levels of each channel were being cv’d by other channels of Batumi and envelops that were being triggered by other channels of Little Nerd, also in Euclidean mode. There was lots of cross modulation between Batumi, uLFO, envelops and probably had some wogglebug in there too.
A relatively simple patch. At times it sounds like a loop, but there’s very subtle, slow moving variation thanks to the very subtle cross modulation. This quality of repetition and super slow moving change is a quality I really like in drone/ambient music.
One main section with 5 or 6 of the SMR channels sounding, then embellished with two additional runs through the same patch, but with slightly different settings in rate on the LFO controlling the whole thing and just one or two deliberately selected notes sounding on the SMR.
After reading the discussion way back in the thread about stereo spread and keeping bass centered, I decided to try out some of the suggestions and use a M/S eq plugin on the various tracks; reinforcing the bass on the M channel and higher register tones on the S channels. This really changed the track for me and brought some clarity and space to it, so thanks for the suggestions in that discussion! I found it really helpful when working on this.
Great track and write up!
Starting my trek through rhe pieces…
Played yours twice. Nice description of the process and place deepers my appreciation.
Directions Track Notes:
Echoes of echoes, memories of memories… thinking over and over about the thing you did or didn’t do, said or didn’t say, as if you could do, or not do, or say, or not say it this time around.
Directions stays in this space, stretching and exploring loops of loops of loops of a few seconds of sample material.
The sound source is three short guitar samples played using norns mlr. The output was then looped again (and again), using w/// and magneto, with additional processing from fumana and rings. Finally, everything was stretched (plus delay and reverb) in live.
Nice - is it your work?
Looking it up I also found: http://yokai.com/category/tsukimono/ (The online database of Japanese ghosts and monsters).
Such a pleasing resource…
Not me - just a favorite from back a bit. My comment was kinda tongue in cheek, but I also count among my favorite computer musicians the “independent researchers” as Trevor Wishart would say… tinkerers, fussabouts… where the computer is just a means & all paths are valid as they are in service of some musical goal…
Edit, so three cheers for imperfect computing