Super Piano

SOTA Music AI Implementations Repo

Built on Google XLNet/Transformer-XL SOTA AI Architecture)

Absolutely fantastic SOTA Google Colab (Jupyter) Notebooks to easily and quickly train a SOTA Music AI model and for generating music with Transformer technology (Google XLNet/Transformer-XL)

Huge thanks goes to creators of the original repos/code that made these amazing Notebooks possible :) Thank you very much and the credit is all yours :)

Description of each Super Piano Colab

[THE BEST] Super Chamber Piano (Non-Transformer, pure LSTM, go figure :)

Absolutely fantastic implementation of Chamber Piano that is capable of playing two instruments at once (Piano and Violin). Plays incredibly well and is capable of smooth continuations. Works well with Jazz as well. Check out screenshots, pre-trained model, and the dataset.

Enjoy :)

1) Super Piano Original

This All-In-One Composition Piano Colab/Model that is capable of either generating a perfomance or a pretty decent continuation of your own composition (Custom MIDI continuation option).

Recommended for every purpose as it can handle any Piano composition task.

It also uses block continuation schema, meaning that you can continue your composition to any desired length and you have full control over each continuation block (size, temp, etc).

Most capable of my Super Piano Music AI models :)

A better pre-trained model for this Piano will be coming very soon.

2) Super Piano 2

This is by far the most capable of my Performance Super Pianos. Capable of super fast piano perfomances generation (any lenghs, any number of perfomances/batches).

This version also has a histogram-based control option that produces very nice results.

Does not work well with MAESTRO DataSet so please use provided Super Piano 2 MIDIs DataSet or use your own.

Recommended use: Batch/Monster Perfomance generation based on a Histogram primer key. See instructions at the Generation Section in Colab where I provide examples.

3) Super Piano 3 [DOES NOT PLAY WELL DUE TO PROBLEMS WITH MIDI PROCESSOR and with some other stuff]

SOTA Perfomance Colab/Model.

This Piano also pretty much a very close copy (or rather a reproduction) of the Google Piano Trainsformer Paper/Model/Colab. Definitely check it out so that you can compare the output and reproducability with my implementation.

Custom MIDI options also present but since it is a Performance model, it is not really designed to be controled or influenced.

4) Super Piano 4

Fully Multi-instrumental LSTM Colab based on Google Transformer-XL architecture. Works well. Check it out.

Some stats:

Architecture Google XLNet/Transformer-XL (FastAI/PyTorch)

You are welcome to train original models further or fine-tune them with your (smaller) MIDI dataset.

Thank you and enjoy these amazing Music AI Notebooks :)

Alex