Nowadays, deep machine learning seems to be playing a significant role in many aspects of almost every field like image detectors, computer vision, medical and business analytic. But a multitude of the available models has been designed for image processing/ classification or text-processing where mostly a variant of CNN or feed-forward neural network is used to extract more in-depth features. However, there are many other realistic problems, where further research need to be carried out on deep machine learning models to address some prominent issues. One of these domains (which require further consideration) is that of music generation. Traditionally, music was treated as an analogue signal and was generated manually. However, in recent years, music is conspicuous to technology that can generate a suite of music automatically without any human intervention. To accomplish this task, we need to overcome some technical challenges to successfully generate a composition of notes having continuity and unity. This is where we can try for different Deep learning frameworks on sequential inputs other than images.
For the dataset we have used used piano music from the soundtrack of Final fantasy as it seems to have distinct melodies that is present in majority of the music pieces. We are using MIDI dataset here as the other datasets of music like the mp3, au etc are raw audio files with a lot of noise which messes up the pre process. For examining the data we are going to be using here, Music21 a python toolkit which is used to get the musical notations of a MIDI file. To make our own midi files in the final music creation process it helps us by creating the Note and Chord objects from the dataset here. The Note object consists of : 1)Pitch - frequency of the sound represented with the letters [A, B, C, D, E, F, G] 2)Octave - the set of pitches which is used on the piano here 3)Offset - the location of the note in the piece. The Chord objects behaves as a container for the set of notes that is being played at a given time.
The picture of the LSTM can be seen below
We have trained the networks for 40 and 100 epochs respectively and we can compare the output of both the models performance. To see the evaluation of a model, we can compare the training and validation accuracy as well as the loss of the both the model and to see how the models has performed in the training. From the graph plotted which is mentioned, we can see that both the training and the validation accuracy has a significant increase over the 20 epochs and after that, the validation accuracy have remained constant But the training accuracy have increased to more than 90 percentage after the 50 epochs and we have achieved maximum of 97.44% of training accuracy. When it comes to the training and the validation loss, training loss have showed a significant decrease and remained constant without increasing over the course of the 100 epochs but the validation loss have increases little bit when we compare to the training loss.