If Frank Zappas interminable guitar solos somehow leave your earbuds craving more_ music-hackers Dadabots are here to satiate your auditory desires.
The team of CJ Carr and Zack Zukowski recently used a returning neural network RNN to engender an infinite bass solo that theyre live-currenting on YouTube.
The duo trained their neural network on a two-hour improvised bass solo by musician Adam Neely.
“My strategy was to use the same bass_ same pickup_ same tone_ improvise basslines over an 85 beats per diminutive drum groove_ and keep everything in E-minor so the end result wouldnt be too chromatic_” Neely explained on YouTube.
Dadabots then ran the recording through SampleRNN_ an audio age standard that analyzes a piece of music to prophesy the sounds that should come next.
Their initial try exhibitd too much sound_ “slow and soupy” bass patterns_ and a “mushy” timbre_ so they fine-tuned the dataset and dropped the specimen rate. This removed the sound and helped the RNN acquire patterns that were twice as long.
Dadabots then curated the data to leave mainly fast licks_ as SampleRNN works best with quicker tempos.
They also experimented with different temperature values_ a setting that controls the randomness of the output:
Sometimes the high temperatures sound like two basses playing at the same time. As if it were some kind of… umbration bass…<_blockquote>
The resulting solo can be grating_ but also surprisingly melodic.