360 hr Librispeech training for neural network

Hi, I want to train librispeech which contains variable length of samples.
How to use Dataloader for this ( should I Cut it into framewise and then train) or is their any methods for loading different sample size.
If I follow to cut it into frames wise, its a huge number files which it able to handle properly by the GPU.
Is their any method to write a customized way to train the neural network.