Successive layers transform information. Earlier layers can identify simpler patterns and later ones can combine them for more complex tasks.
A deeper network can represent some structures efficiently, while performance still depends on architecture and training data.
ELI5
For example, an image model can learn edges and shapes before using combinations of them to identify objects.
It learns useful patterns from examples rather than requiring a person to write out every visual clue.
Does more depth always help?
Benefits depend on data and design. Adding layers can also increase the amount of work to train or run the model.
Can it learn from sound?
It can work with text, images, sound or other inputs that the network is trained to interpret.

