Music is mainly an artistic act of inspired creation and is unlike some of the traditional math problems. Music cannot be solved by a simple set of formulae. The most interesting and challenging part is producing unique music without infringing the copyright. The generated music has to sound good, and what sounds good is very subjective and varies from culture to culture.
Artificial Neural Network/Deep Learning has a wide range of applications, such as in Image processing, Natural language processing, Time series prediction, etc. But what about its usage in art? Could we use deep learning to create music?
This talk is about how deep learning models were used to produce music - catering particularly to Bollywood.
This talk would show how an exquisite piece of art i.e. music can be generated using deep learning model which helps in automated feature extraction. In order to automate the music generation, the model must be able to remember the learned features over the longer period of time, this is achieved by a special type of Recurrent Neural Network (RNN) called as LSTM (Long Short Term Memory) network.
Implementation of such complex model can be made much easier using inbuilt Python libraries such as Keras with Theano as backend. It allows for easy and fast prototyping. Packages like numpy and scipy are being used for easier mathematical computation of input vectors and for reading/writing the WAV files respectively. The neural network architecture makes use of numerous amount of music samples to train the model. After an adequate number of iterations and training time, this model generates music that is unique and original.
In this talk, steps involved in preprocessing of data, training the model, testing the model and generating the music from the trained model will be discussed. This talk will also cover some of the challenges and tradeoffs made for algorithmic music generation.