Bergiza Bot

DJ setlist generator using Spotify audio features data.

Link to setlists: https://soundcloud.com/bergizabot/sets

Installation

pip install -r requirements.txt

Generating set lists

Spotify API

The SetListGenerator_Spotify class is the primary component to drive track generation. It can be found within 'SetlistGenerator_SpotifyAPI.ipynb'

d = SetlistGenerator_Spotify(seed_tracks=[list_of_track_ids],
                             seed_genres=[list_of_genres],
                             seed_artists=[list_of_artist_ids])

When running the script for the first time, a webpage will be opened under the localhost domain. Copy and paste the URL into the prompt to authenticate with the Spotify API

Ex. 1 Generate a 10-track set from Aphex Twin's track 'Donkey Rhubarb'

d = SetlistGenerator_Spotify(seed_tracks=['0c0PhGTdl4K1SRWu0kYkBh'],
                             seed_genres=['minimal-techno', 'techno'])

while len(d.set_list) < 10:
   d.getNextTrack()
   d.changeSeedTracks()
   d.expandTrackPool()

d.set_list

In order for the generator to work properly, all seed_ arguments passed must be in the form of a list.

To retrieve a single artist id: search_for_artistID(search_string)
To retrieve a single track id: search_for_trackID(search_string)

(When searching for a track, be sure to use proper spelling and include the artist's name to ensure it retrieves the correct track)

Current Issues

Some of the following bugs have been identified:

Using more than 2 seed_genres at a time can produce errors
Using more than 2 seed_artists at a time can produce errors
Using more than 3 seed_tracks at a time can produce errors

Using Recurrent Neural Networks

Located within 'lai_rnn_narrative' there is experimental code to produce set lists with RNNs. .ipynb files are codes for:

parsing tracklist and feature vectors,
training narrative model of vectors by RNN(LSTM) using keras
generating tracklist by predict vectors using trained model and select nearest neighbor from random samples (and some testing blocks)

Currently all_rnn.ipynb is not working. .h5 files are the trained model. Training session can be skipped by calling model.load block and not calling training blocks some generated example can be found in .txt

Other methods

As of V_0_0_9, the SetListGenerator_Spotify generator does not utilize mass data scraped from MixesDB, Discogs, or Spotify. However, the experimental code in 'lai_rnn_narrative' does use this data. While the results are still somewhat opaque in their selection criteria, this section of the system is still valuable and thus has been included in the current version.

iamthevastidledhitchhiker / DJSetListGenerator