If you have participated in a fight in the last two weeks you may have noticed some different sounding pieces being submitted.
Last week: Take It Outside - Deep Orb Whisky - http://www.songfight.org/music/take_it_ ... ky_tio.mp3
This week: I Am Very Busy - Deep Orb Steel - http://www.songfight.org/music/i_am_ver ... l_iavb.mp3
These songs were generated using home grown RNN Based learning techniques as described here https://towardsdatascience.com/how-to-g ... 786834d4c5
Right now there are three models trained.
1. Deep Orb Whisky - Training from over 200 public domain midi country, western and bluegrass songs. Lyric generation from public domain cowboy poetry from 1891-1954
2. Deep Orb Steel - Training from over 170 public domain metal, speedmetal and dark metal songs. Lyric generation from public domain gothic poetry including Poe, and HP Lovecraft
3. Deep Orb Sparkle - Training from over 290 public domain pop, electro, and EDM songs. Lyric generation from public domain young adult poetry from 1945-2000
How does a collab work?
In order to maintain the spirit of the collab I am proposing some basic rules (with artistic license)...
- Humans choose the model Whisky, Steel or Sparkle
- Humans receive midi generated by DeepOrb. This is a multitrack "song"
- Humans can cut, copy, paste anything down to the unit of one measure for AI-generated tracks. Strange outlier notes can be deleted but an effort to retain the vast majority of the measure should be made.
- Humans can produce up to 50% of the measures of the song. That might mean accompanying tracks, it may mean inserting MIDI in an AI generated track.
- Humans can assign any instrument or effect they wish to any track.
- Humans may choose to sing or can have MP3 lyrics generated for them, but should try to use artistic licence with the generated text to make an interesting set of lyrics with at least 50% of lyrics coming from the AI.
* All collaborations become public domain and will be put in song play rotation with attribution on https://deeporb.net