Robust Audio Identification for Commercial Applications

Along with investigating similarity metrics between audio material, the topic of robust matching of pairs of audio content has gained wide interest recently. In particular, if this matching process is carried out using a compact representation of the audio content ("audio fingerprint"), it is possible to identify unknown audio material by means of matching it to a database with the fingerprints of registered works. This paper presents a system for reliable, fast and robust identification of audio material which can be run on the resources provided by today's standard computing platforms. The system is based on a general pattern recognition paradigm and exploits low level signal features standardized within the MPEG-7 framework, thus enabling interoperability on a world-wide scale. Compared to similar systems, particular attention is given to issues of robustness with respect to common signal distortions, i.e. recognition performance for processed/modified audio signals. The system's current performance figures are benchmarked for a range of real-world signal distortions, including low bitrate coding and transmission over an acoustic channel. A number of interesting applications are discussed.