A robust audio fingerprinting method using spectrograms saliency maps

Audio fingerprinting is widely used for audio identification, indexing, searching, navigation, monitoring and other monetization purposes, as well as support to other areas such as watermarking, music information retrieval and video identification. Because the ease of distorting intentionally or unintentionally an audio signal, the robustness and accuracy are very important characteristics in audio fingerprinting to avoid false alarm errors. In this paper, an audio fingerprinting method is proposed, it uses the spectrogram representation of an audio signal, combined with a global fingerprint extraction. The fingerprint extraction is based on the saliency maps of the spectrogram of the audio signal. The proposed method is robust against common signal processing attacks, colored noise addition and severe temporal desynchronization attacks, e.g. time shifting, pitch shifting and speed change.