Extracting Features from Audio Samples for Machine Learning

Humans are great at classifying noises. We can hear a chirp and surmise that it belongs to a bird, we can hear an abstract noise and classify it as speech with a particular meaning and definition. This relationship between humans and audio classification forms the basis of speech and human communication as a whole. Translating this incredible ability to computers on the other hand can be a difficult challenge, to say the least. Whilst we can naturally decompose signals, how do we teach computers to do this, and how do we show what parts of the signal matter and what parts of the signal are irrelevant or noisy?