Generate lyrics in the style of your favorite artists 🎶
artist2lyrics is a set of CLI tools that lets you create your own lyrics generator AI, simply by specifiying the artists whose lyrics style you want your AI to imitate. Lyrics crawling, pre-processing, word embeddings generation, model training, and lyrics generation - all basic steps of the lyrics generation pipeline are already handled by this tool! By default, artist2lyrics uses a 2-layer stacked LSTM RNN with dropout to learn lyrics in a word-based manner.
Well then, let's see how this works.
To start off, I'll be using music from my country. Using artist2lyrics, I trained a model on Tagalog songs of three of my loved Filipino bands: SugarFree, Itchyworms, and Eraserheads. Now, generating some lyrics...
barkada . hawak ang 'yong kamay habang ay hindi . naaalala ka pa .
ilaw ikaw sarili ng mundo ng ilalim ng buwan ilalim ng buwan .
Hmmm... so this AI seems to be thinking about his friends, a loved one, and the moon.
Let's generate some more.
bituin sa ating pugad , sa ating pugad sa ating pugad , sa ating pugad .
bawat sandali ng aking buhay puno mo ang aking kamay .
OK that's pretty sweet. This AI's feeling at home with a loved one.
. kahapon lang tayo ay buhay ay yakapin , ako ay buhay buhay ako
hay buhay nga naman hay buhay nga naman .
Now it's frustrated about life.
Let's try something else.
After training a model on more than 150 songs of Air Supply, we get these:
see stand it the feeling when you're far away
i want to take you up in my little world so you'll be here today .
when we fly through the night can't you feel the words is the truth of the line .
let what being here with you taking what is mine just between the tears .
i could not more cry i say that i've i thought in heart alright
but that heart is a face , can't you thought know just tonight i wanna survive you
Well this AI's kinda weird and funny, but surely it wants to sing about love.
Let's try a different genre.
right after me you want to kick up you're dead ? .
so play a last to remember to the head . i will want .
so maybe a sad song with nothin' to do
about a life long wait for a hospital about a life
about well better know get a our way it's .
good is you , you , you . me and you and all of these living dead ,
burning up in the sun where the bodies go , i've on , and on ,
and on , and on , and on . so the red your - that what's inside
you'll a worst from have out gone .
Great. I have just raised an emo kid AI. 😆
Okay, generating lyrics sure is fun. Now let's get to more FUN!
- PostgreSQL
- Python3
- numpy, tensorflow, keras, gensim, matplotlib, psycopg2, scikit_learn
- see requirements.txt for specific versions used
By default, artist2lyrics uses postgres database to store the crawled lyrics.
To be able to use this database, you need to register your user account as a user:
sudo -u postgres createuser <user>
The pipeline for this tool's lyrics generation is as follows:
Lyrics Crawling -> Pre-processing -> Embeddings Generation -> Model Training -> Lyrics Generation
Each of these are handled by a particular Python script. Before running these, make sure to set desired configuration in artist2lyrics.cfg. Importantly, to specify which artists you'd want to get lyrics from, set the "artists" field in the config file
in comma separated format. For example:
artists = airsupply,mychemicalromance
Note that for lyrics crawling, we are gathering data from azlyrics. The specified names of the artists shall match the name appearing in azlyrics url. For example:
https://www.azlyrics.com/m/mychemicalromance.html
https://www.azlyrics.com/a/airsupply.html
With these set, you may proceed with execution of the scripts in the following sequence:
- crawl_lyrics.py - lyrics crawling
- process_data.py - data pre-processing and embeddings generation
- train_model.py - model training
- generate_lyrics.py - use trained model for lyrics generation
This project is licensed under the Apache License, Version 2.0 - see the LICENSE file for details
Credits to
- IIITV org for the lyrics-crawler
- Kyubyong for word2vec demo
- Jeff Delaney for t-SNE demo