Sound-mnist model training performs much worse on own computer, compared to Github

Question

I just pulled this GitHub repository which builds a CNN for MNIST sound classification. It is straight forward to use, I just cloned it, made sure I had all the libraries in