Abstract: By means of Gammachirp-cochleagram, this paper presents a novel time-frequency representation algorithm for audio features. The local feature of the Gammachirp-cochleagram is firstly extracted by non-negative matrix factorization (NMF). And then, both the difference and the quantization are applied on the extracted features to further enhance its robustness and reduce its computational complexity. Experimental results illustrate that the proposed algorithm achieves superior performance in robustness and identification rate under the attack of audio editing software and the searching of record.