Why machines struggle to understand speech

source