Detecting Subtle Differences between Human and Model Languages Using Spectrum of Relative Likelihood