Enhancing speaker recognition with multiple observations over mobile networks