Comparison of Algorithms for Audio Fingerprinting

Heinrich A. van Nieuwenhuizen, Willie C. Venter and Leenta M.J. Grobler

School of Electrical, Electronic and Computer Engineering
North-West University, Potchefstroom Campus, South Africa
Email: {20252188; willie.venter; leenta.grobler}@nwu.ac.za
Telephone: (027) 18 299-4058 Fax: (027) 18 299-1977

Abstract-The practical implementation and application

Recordings’
of audio fingerprinting to recognize specific audio collection
Fingerprint
segments are becoming more and more popular these extraction

days. Different research groups are working on

Recordings’
implementations of audio fingerprinting, but they IDs
seldom, compare their algorithms to those of other DB

people working in the field. A fair judgment can

therefore not be made of which of the available
algorithms is suitable for certain applications. In this
paper, two audio fingerprinting algorithms are tested Unlabeled Recording
recording ID
that of Avery Wang’s and Haitsma and Kalker’s in Fingerprint
Matc
extraction
terms of accuracy, speed, versatility and scalability

Keywords: Audio Fingerprinting; Automatic Music Recognition;

Content-based Audio Identification; Perceptual Hashing; Robust Figure 1 : Content-based audio identification framework [6]
Matching;
In section II, the three main groups of audio
fingerprinting techniques are briefly discussed, after which
I. INTRODUCTION
the operation of Avery Wang’s Shazam [2] and Haitsma and
Fingerprinting systems are not a new concept; it has been Kalker [1] algorithm are briefly discussed. In section III the
around for more than a hundred years. In 1893 Sir Francis operation of the algorithms are discussed. In Section IV the
Galton was the first to “prove” that no two fingerprints of validation and verification is presented. The paper is ended
human beings are alike [1]. This notion was taken further by with a conclusion and future work in section V.
using any unique feature to identify an object; this includes
the iris and even ears. People also realized the potential of II. DIFFERENT AUDIO FINGERPRINTING TECHNIQUES
constructing fingerprints of audio signals to identify and
compare them. This principle is called audio fingerprinting. There are several applications for audio fingerprinting
algorithms. According to Wes Hatch [3] the biggest
Audio fingerprints make use of short audio segments benefactor would be the broadcast monitoring industry.
usually between 3-30 seconds in length (depending on the Other applications would be playlist generation, royalty
algorithm) to create an audio fingerprint. This audio collection, program verification and advertisement
fingerprint is compared to a database of known audio verification.
fingerprints to identify the original audio source. The audio According to P.J.O Doets, M. Menor Gisbert and R.L.
fingerprints of the segments do not necessarily have to be of Lagendijk there are three groups [4] which audio
high quality to be a match. Distortions and interference of fingerprinting can be classified as:
the original signal makes matching of the fingerprints less
reliable, but (to a certain extent) it will still be recognizable. Group 1: Systems that use features based on multiple
The distortions and interferences can be compared to a subbands, namely Philips’ Robust Hash algorithm, reported
smudged or partial human fingerprint. to be very robust [1] against distortions. Phillips uses
Haitsma and Kalker’s algorithm.
In this paper a brief description about audio
fingerprinting, a generic code of Avery Wang’s Shazam Group 2: Systems that use features based on a single band
algorithm and that of Haitsma and Kalker’s algorithm are such as the spectral domain, namely Avery Wang’s Shazam
discussed [2]. and Fraunhofer’s AudioID algorithms.

Group 3: Systems using a combination of subbands or

frames, which is optimized through training, namely
Microsoft’s Robust Audio Recognition Engine (RARE) that
uses Hidden Markov Models (HMMs). [5]

This work was completed at the Telkom-Grintek Centre of Excellence

At the NWU, and is funded by the HTBO THRIP project
 III. OPERATION Landmark-Based Audio Fingerprinting” [5]. Verifying that
the algorithms are correctly implemented.
The Haitsma and Kalker’s algorithm propose that a The two algorithms are then compared to each other in
terms of accuracy, speed, versatility and scalability. The
fingerprint extraction scheme should be based on a general
algorithms are further tested by subjecting the data to noise
streaming approach, taking an audio signal framing it into
and compression and the algorithms parameters are tweaked
windows of 370ms length for every 11.6ms giving it an
for better results. Comparing the algorithms in the latter
overlapping factor of 31/32 [1]. terms but with different data ranging from classical music to
pop and advertisements.
After which the FFT of every frame is computed filter
though band division stored into 32-bit sub Fingerprints. V. CONCLUSION AND FUTURE WORK
See figure 2 below
The following were observed in the study.
Increasing and decreasing both algorithms frame size will
decrease speed but increase the accuracy respectfully.
Defining more peaks in the Shazam’s algorithm’s frame
(normally 5) would also result in better accuracy but
decrease speed. Increasing the overlapping regions in both
algorithms will increase robustness but decrease speed.
The nearest false positive is on average 5 or 6, which
strengthens James .P Ogle and Daniel P.W. Ellis theory that
9 peaks[7] are needed to be a match.
In the future the following should be tried
Figure 2 : Haitsma and Kalker’s algorithm [1] Translating the algorithms to a faster programing
environment. Further research should be done on different
An unidentified sample of an audio signal should be 3-30 techniques to access the database quicker.
seconds in length for their algorithm to identify a match. More application uses should be investigated e.g.
Avery Wang claims for a database of 20 thousand tracks gunshots, engine noise etc.
implemented on a PC, the search time is 5 to 500
milliseconds [2]. As the code is not available, generic code
VI. BIBLIOGRAPHY
for MATLAB was produced by Dan Ellis [5]. Robert Macrae
of C4DM Queen Mary University London altered the code [1] J. Haitsma and A. Kalker, "A Highly Robust Audio Fingerprinting
for use in the windows environment. System," International Symposium on Music Information Retrieval
(ISMIR), pp. 107-115, 2002.
The latter algorithm proposed to make use of a [2] A. L.-C. Wang, "An Industrial-Strength Audio Search Algorithm,"
spectrogram. The spectrogram is the squared magnitude of ISMIR, 2003.
the STFT (Short-time Fourier Transform). [3] W. Hatch, "A Quick Review of Audio Fingerprinting," Mar. 2003.

spectrogram(t ,  )  STFT (t ,  )
2 [4] P. J. O. Doets, M. M. Gisbert, and R. L. Lagendijk, "On the
(1) comparison of audio fingerprints for extracting quality parameters of
compressed audio," vol. 6072, 2006.
Usually the spectrogram is divided into small sizes [5] D. P. W. Ellis. (2009) Robust Landmark-Based Audio Fingerprinting.
(typically 512 points) which are called windows or frames. http://labrosa.ee.columbia.edu/matlab/fingerprint/
This is the shared basis of group 2. The differences between [6] P. Cano, E. Batlle, E. Gómez, L. de C.T.Gomes, and M. Bonnet,
the fingerprint algorithms in the group typically involve "Audio Fingerprinting: Concepts And Applications," Studies in
Computational Intelligence (SCI), no. 2, pp. 233-245, 2005.
how much the frames overlap, and how the fingerprint is
[7] J. P. Ogle and D. P. W. Ellis, "Fingerprinting to identify repeated
defined in the frame and the storing and searching of the sound events in long-duration personal audio recordings," 2007.
fingerprints. Avery Wang Shazam algorithm uses the energy
peaks in the frame and form spectral pair landmarks. The
local maxima within a defined section are grouped into pairs Heinrich van Nieuwenhuizen received his B.Eng degree in 2009 and is
[4]. currently pursuing his M.Eng at the North West University, Potchefstroom
The hash values are computed and compared the entry campus. His research interests include software design, audio fingerprinting
and implementation and comparison of audio fingerprinting algorithms for
with the most hits is returned as the match. (Typically more industrial use.
than 9 spectral peaks are considered a match [7].)

IV. VALIDATION AND VERIFICATION

The two algorithms are implemented. Haitsma and
Kalker’s results are compared to that of their own “A
Highly Robust Audio Fingerprinting System” [1] and Avery
Wang’s compared to his own “An Industrial-Strength Audio
Search Algorithm” [2] and that of Dan Ellis “Robust

This work was completed at the Telkom-Grintek Centre of Excellence

At the NWU, and is funded by the HTBO THRIP project