For　the　popular　audio　bandwidth　extension,　the　stationary　signal　processing　and　spectral　patching　methods　are　utilized　to　artificially　recover　high　frequency　components,　yet　it　inevitably　leads　to　the　undesired　auditory　roughness.　In　this　paper,　a　nonlinear　bandwidth　extension　method　of　audio　from　wideband　to　super-wideband　was　proposed.　Phase　space　is　reconstructed　from　the　fine　structure　information　of　audio　spectrum　and　the　trajectory　of　phase　points　is　efficiently　predicted　with　nearest-neighbor　matching.　In　addition,　the　spectral　energy　of　high　frequency　is　estimated　based　on　Gaussian　mixture　model　(GMM).　Both　the　objective　and　subjective　test　results　demonstrate　that　the　proposed　algorithm　has　better　performance　than　conventional　blind　bandwidth　extension　algorithms.