Music Aids

Trained Audio File

4 Things to know about our Music Aids

1. Modern music sucks? Make your own. No inspiration for music? No problem. Tone-deaf? Why worry man.
2. User can record and upload any audio to our telegram bot and we will compose a new music based on their recording
3. Based on recurrent neural network to provide better music/melodies for the user.
4. Can compose simple notes, piano pieces, play drums, or a mixed of everything. (Except for Justin Bieber's voice)

Problem Statement/Opportunity

The music industry is a $3 billion industry, yet what it have done in the past decade is generate poison for our ears. !

A Spanish research agency have proven that in the past few decades pitch transitions have decrease, the homogenization of the timbrel palette, and growing loudness levels in music. (Serrà, Corral, Boguñá, Haro und Arcos 2012).

This bot works perfectly well as a standalone bot but it can still be the perfect tool for composers to obtain inspiration.

Lastly, as much as we understand that artificial intelligence is no match for natural stupidity of 'musicians' we still believe in striving selfless and noble goal. P.S. Most Socially Useful or Environmental Friendly Hack exist this year.

Solution Outline

User input desired model to use and upload necessary files
Bot will run the relevant model according to the inputs
And then magically a musical masterpiece is created and sent back to the user

Technology Used

Implement a myriad of machine learning models in our bot. Technology such as RNN, LSTM, RNN-NADE, RL, VAE and generative model. Models from multiple machine learning model from multiple well known library such as Magenta and Tensor-flow are implemented. For example, we used Drums RNN, Melody RNN, Music VAE (a recurrent network and autoencoder for music). We also used the ouput of one model as the input of another model to improve the accuracy.

(Good luck Installing them)

Difficulty faced

There are no simple solution to conversion of audio files. (Read more to understand how cancerous it truly is)
- A simple conversion of a mp3 to MIDI file (Default input for audio machine learning) will require installation of 7 dependencies packages. 2 of which are unable to run in python and a work around using command line have to be used. Conversion using python3 and above is not working and python2.7 have to be installed instead.
- The reverse conversion is not any better.
Music taste is subjective and not quantifiable which makes it hard to train the model.
It is also an unexplored region where many projects are none open sourced and training data is hard to obtain.
None of us have any experience with machine learning in music to worsen it the current repository of machine learning is none user friendly. Most are bash code with no explanation or proper explanation in python.
Telegram Bot server is down from 6am till 8:01:25am