In this paper, we present the latest improvements on spectrogram-matrix based fingerprinting system for detecting transformed audio copies. In particular, we experiment with two feature parameters derived using global and local spectrogram averages and show that combining results from these two feature parameters significantly improves performance. We test our system on TRECVID 2010 content-based copy detection dataset. Experimental results show the robustness of our method against various audio distortions. The proposed method reduces the minimal Normalized Detection Cost Rate (min NDCR) by 23% and improves localization accuracy by 24% compared to a state-of-the-art audio fingerprinting system. Our system achieves the lowest min NDCR for MP3 compression transformation with a relative improvement of 60%.
Bibliographic reference. Ouali, Chahid / Dumouchel, Pierre / Gupta, Vishwa (2014): "Robust features for content-based audio copy detection", In INTERSPEECH-2014, 2395-2399.