My Expedition to Machine-Generated Music Using Tensorflow Magenta Part 1: Linkin Park

2 minute read

Repository

https://github.com/tensorflow/magenta

Introduction

Linkin Park fans cannot forget their music. Sadly, we cannot hear it again, because of the death of the late Chester Bennington. Although replicating his voice is unfortunately impossible based on current technology, replicating all the other instruments are possible.
I have tried using two NVIDIA Tesla V100 16GB SXM2 GPUs and the MusicVAE model (https://magenta.tensorflow.org/music-vae) to generate music based on MIDI files of songs from the albums of Linkin Park.

Post Body

image.png Can we forget him? No. (Image courtesy of Linkin Park) image.png Can we forget them? No. (Image courtesy of Linkin Park)

I became a fan of Linkin Park a bit before his death. That was when I was starting to listen to a lot of music. I didn’t before. And it was shocking to see him go months into Linkin Park.

Because of this, when I found out about the Magenta project while I was working on a programming project related to music, I thought instantly of Linkin Park. If I could replicate their Nu-metal music using Magenta.

I took no time wasted and immediately trained around 15 Linkin Park Midi files, composed of songs that are in the nu-metal albums Hybrid Theory and Meteora. This Link is the downloadable .tfrecord file generated using This Manual.

Refer to Training Your Own MusicVAE for how you can train yourself. I have used the hierdec-trio_16bar config and you will want to use the same if you are working on replicating popular music or band music like me. You can also generate piano music like Bach with the Performance RNN Model in the same package.

It took me about a week on one Tesla V100 GPU for 200k steps, in a batch size of 32.

Generation was done using This Command. Five samples were generated.

image.png image.png image.png image.png image.png

The results were not impressive. Because of overfitting, the sounds resembled one specific song in one of the Linkin Park’s nu-metal albums. The results are available Here. (Quiz: What is the song?)

In my next post, I will address the expansion of data trained to create actual original music that replicates Linkin Park. This one is gonna be actually good. Please look forward to it!

Although I think my first test was unsuccessful, please rate the generated music and how you think about it!

If you want to generate new music using this trained model, Download the Checkpoint from Here and use Tensorflow Magenta to generate using This Command.

Resources

MusicVAE model (https://magenta.tensorflow.org/music-vae) in Tensorflow Magenta (https://github.com/tensorflow/magenta)
FreeMidi.org (https://freemidi.org/)
FLAC Synthesizer by Musescore (https://musescore.com/ or https://github.com/musescore/MuseScore)

This is the first post of a series of blog posts on Machine-Generated Music. The collection of posts will expand based on new selections on genres and music artists or bands.

Second Post

Comments