Statistical Analysis of Randomness in Training of Small-Scale Neural Networks for Speech Enhancement

For the training of neural networks, randomness due to the initialization is a widely acknowledged factor for training success. In speech enhancement, randomness is also introduced via the creation of the training dataset due to the synthetic mixing of speech and noise, and other data augmentation techniques. This aspect of randomness is often neglected in literature. In this contribution, we quantify the effect of randomness in the generation of the training dataset in comparison to the randomness introduced by the network initialization and find that for small networks and very small datasets, the assembly of the training dataset has a significant influence on the network’s performance.