Abstract:
Audio fingerprinting is a smart technology to
identify music contents relevant to the query by
comparing the content-based hash (fingerprint) of the
query to known hashes in the fingerprint database. For
a million-song library, size of the fingerprint database
restricts speedy and correct music identification. In
this paper, we present a space-saving audio
fingerprinting system based on Mel Frequency
Cepstral Coefficients and analyze in detail. For a
number of cepstral coefficients with 12 as default, the
system yields a 2712-bit fingerprint for a 3-sec audio
clip, a significant reduction in storage compared to the
8192-bit fingerprint of well-known Philips Robust
Hashing method. Experimental results also show that
the more number of input value for cepstral
coefficients, the larger the fingerprint size. The choice
of Mel coefficients also affects the identification
performance of the system: 8 to 12 coefficients gives
the best similarity rates while preserving robustness of
the fingerprints to signal distortion such as pitch
shifting.
