Good day everyone.
I am engaged in machine learning and recently switched to RNN. I'm trying to solve the problem of generating music, but not in the form of a midi file, but as a set of PMC data. As far as I understand from all the articles I have read, it is the recurrent networks that can cope with this task. The question is that I do not understand how it is necessary to set the input_shape model so that it learns. At the input I have a one-dimensional array of converted audio file data with a given samplerate. I would like to get a grid at the output, which would continue to generate sound when a random signal was applied to the input.
Having sat with this problem for about a week, I didn’t figure out how to make the resulting model unfold over the entire length of the input data and continue to generate the sequence of the desired length. Each time the models that I received were trained only on the allocated number of the last points in the data sequence, and did not perceive the entire audio file, which is the main advantage of the RNN, as I understand it.
Python3, tensorflow, keras is used.
