A calibration of various microphones that have different characteristics is very difficult. This paper presents a feature extraction method as an alternative. The method provides acoustic features that are strongly robust against various characteristic transfer functions. The proposed method applies Local Binary Patterns (LBP) and Compressive Sensing (CS) which compare spectral details with spectral envelopes of their spectra. A visualizing high-dimentional data analysis method, t-SNE, is performed as a verification. The results show that a classification using the proposed feature is very impressive. Multiple classes with various microphones are grouped correctly.