Congratulations! You have just found DeepSpotify.
DeepSpotify learns your favorite Spotify playlists to generate music. It uses Melodia and spotipy/Spotify Web API to extract a melody from your favorite Spotify playlist, Keras to learn the grammars of these melodies with a two layer LSTM RNN, and generates a new melody that should sound like the other songs in that playlist.
Follow my progress on this project on my blog, alexisafk.com.
Make sure to have Anaconda and Melodia installed. See further down for instructions on how to install Melodia.
Then create an Anaconda environment from the DeepSpotify.yaml in the project root directory by running:
$ conda env create -f DeepSpotify.yml
$ source activate DeepSpotify
This will create an environment called DeepSpotify and install the necessary libraries that you need to run DeepSpotify.
P.S. Detailed instructions on how to install Melodia 1.0 are here.
Manually install all these dependencies (and their dependencies) in a Python 3.6 environment:
- tensorflow=1.10.0
- Keras==2.2.2
- spotipy==2.4.4
- Melodia==1.0
- vamp==1.1.0
- librosa==0.6.2
- music21==5.3.0
- MIDIUtil==1.2.1
- matplotlib==2.2.3
main.py is run by passing two parameters:
$ python src/main.py -h
Using TensorFlow backend.
usage: main.py [-h] run_opt uri
positional arguments:
run_opt An integer: 1 to analyze audio, 2 to train, 3 to generate, 4 to
do everything.
uri A string: spotify playlist uri
optional arguments:
-h, --help show this help message and exit
The run_opt is used to run different parts of the program. Keep in mind that (1) analyzing audio is required before (2) you can train the model which is required before (3) you can generate new songs. Use (4) to run everything all at once. This will take quite some time to run.
The uri is the Spotify playlist URI. This should in a format like this spotify:user:alezabeth1997:playlist:1LMwOJtfhsnN2FGf0yizXl (an example of my personal Snarky Puppy playlist).
After a complete run of DeepSpotify, the project directory will look like this:
- data/
- melodia/
- {playlist_name}/
- data/
- synth/
- model/
- {playlist_name}/
- new_synth/
- stat/
- {playlist_name}/
data/melodia/{playlist_name}/data/ - timestamp + frequency csv melody data will be placed here.
data/melodia/{playlist_name}/synth/ - for each song, you'll find a:
{song}.orig.wav(a wav of 30s Spotify preview),{song}.melo.and.orig.wav(a wav of extracted melody superimposed on original preview), and a{song}.melo.midi(midi interpretation of extracted melody).
data/model/{playlist_name}/ - after each epoch for each song the model is trained on will be saved in the format model-dX-eX.hdf5.
new_synth/ - generated song called {playlist_name}.mid will be placed here.
stat/ - statistical summaries of model training; includes Tensorflow events to be opened with Tensorboard.
Because all music is not equal, the quality of the generated songs may vastly differ. This is probably attributed to how Melodia can extract the melodic frequencies better from some songs than for other songs (i.e., cleaner melody in a jazz piece than an Eminem rap).
Feel free to fork this project! I'll be making improvements to this project (i.e., considering using raw melodies instead of midi files because midi doesn't seem to capture melody accurately, feeding in model data differently, etc.). Follow my progress on my blog!