Robust audio fingerprinting for song identification

This paper presents a new algorithm for audio hashing. These kinds of techniques work analysing the signal taking into account its nature in order to extract distinguishing features robust to typical processing. These features describe in a compact way the signal and can be efficiently stored in a database. In our algorithm the features are analysed for short frames using a quantization approach. Moreover we propose a parameter for identification, which is also used to define a confidence. The confidence is used to identify the quality of the identification. The algorithm is tested under different processing operations: compression, cropping, noise addition, subsampling, stereo to mono conversion, etc. The results show that the identification can be performed also using a short excerpt of the song.