As a classical audio fingerprint algorithm, the Philips algorithm has been widely used. However, the feature extraction module of the Philips algorithm is inefficient and time-consuming, and the feature binarization algorithm is susceptible to noise interference, which affects the correct rate of audio fingerprint matching. To solve the problems mentioned, this paper proposes an improved Philips audio fingerprint algorithm. The Gammatone filter bank is used to analyze the frequency spectrum of the audio signal to simulate the frequency-selective characteristic of the basilar membrane of the human ear, and a graph-convolutional neural network is introduced to extract the global features between different frequency bands. The distance correlation coefficient is used to calculate the distance between the audio feature matrices to achieve the matching of audio fingerprints. The experimental results show that compared with the original Philips algorithm, the algorithm proposed in this paper achieves lower time consumption and stronger noise immunity.