Arthur Flexer∗ and - IA de Investigación especializada en el MIT

Arthur Flexer∗ and
Dominik Schnitzer∗†
∗Austrian Research Institute for Artificial
Inteligencia
Freyung 6/6
A-1010 Vienna, Austria
arthur.flexer@ofai.at
†Department of Computational
Percepción
Johannes Kepler University Linz
Altenberger Str. 69
A-4040 Linz, Austria
dominik.schnitzer@jku.at

Effects of Album and Artist
Filters in Audio Similarity
Computed for Very Large
Music Databases

In music information retrieval, one of the
central goals is to automatically recommend
music to users based on a query song or query
artist. This can be done using expert knowl-
borde (p.ej., www.pandora.com), social meta-data
(p.ej., www.last.fm), collaborative filtering (p.ej.,
www.amazon.com/mp3), or by extracting informa-
tion directly from the audio (p.ej., www.muffin.com).
In audio-based music recommendation, a well-
known effect is the dominance of songs from the
same artist as the query song in recommendation
liza.

This effect has been studied mainly in the
context of genre-classification experiments. Porque
no ground truth with respect to music similarity
usually exists, genre classification is widely used for
evaluation of music similarity. Each song is labelled
as belonging to a music genre using, p.ej., advice
of a music expert. High genre classification results
indicate good similarity measures. Si, in genre
classification experiments, songs from the same
artist are allowed in both training and test sets, este
can lead to over-optimistic results since usually all
songs from an artist have the same genre label. Él
can be argued that in such a scenario one is doing
artist classification rather than genre classification.
One could even speculate that the specific sound of
an album (mastering and production effects) is being
classified. In Pampalk, Flexer, and Widmer (2005)
the use of a so-called “artist filter” that ensures that
a given artist’s songs are either all in the training set,
or all in the test set, se propone. Those authors found
that the use of such an artist filter can lower the

classification results quite considerably (as much as
de 71 percent down to 27 por ciento, for one of their
music collections). These over-optimistic accuracy
results due to not using an artist filter have been
confirmed in other studies (Flexer 2006; Pampalk
2006). Other results suggest that the use of an artist
filter not only lowers genre classification accuracy
but may also erode the differences in accuracies
between different techniques (Flexer 2007).

All these results were achieved on rather small
bases de datos (de 700 a 15,000 songs). Often whole
albums from an artist were part of the database,
perhaps even more than one. These specifics of
the databases are often unclear and not properly
documented. The present article extends these
results by analyzing a very large data set (encima
250,000 songs) containing multiple albums from
individual artists. We try to answer the following
preguntas:

1. Is there an album and artist effect even in

very large databases?

2. Is the album effect larger than the artist

efecto?

3. What is the influence of database size on
music recommendation and classification?

As will be seen, we find that the artist effect does
exist in very large databases, and the album effect is
bigger than the artist effect.

Datos

Computer Music Journal, 34:3, páginas. 20–28, Caer 2010
C(cid:2) 2010 Instituto de Tecnología de Massachusetts.

For our experiments we used a data set D(ALL)
of S = 254,398 song excerpts (30 seconds each)
from a popular Web store selling music. The freely

Computer Music Journal

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Cifra 1. Percentages
(y-axis) of albums having
1, 2, 3, 4, o 5 a 8 genre
labels (x-axis).

Cifra 2. Percentages
(y-axis) of artists having 6,
7, …, 19, o 20 a 29
albums (x-axis).

5 a 8

Mesa 1. Percentages of Songs Belonging to the
22 Genres, with Multiple Membership
Allowed

Genre

Pop
Classical
Broadway
Soundtracks
Christian/Gospel
New Age
Miscellaneous
Opera/Vocal
Alternative Rock
Rock
Rap/Hip-Hop
R&B
Hard Rock/Metal
Classic Rock
Country
Jazz
Children’s Music
Internacional
Latin Music
Folk
Dance & DJ
Blues

Percentage

49.79
12.89
7.45
1.00
10.20
2.48
6.11
3.24
27.13
51.78
0.98
4.26
15.85
15.95
4.07
6.98
7.78
9.69
0.54
11.18
5.24
11.24

10 11 12 13 14 15 16 17 18 19 >20

parsed automatically from the HTML code. El
excerpts are from U = 18,386 albums from A = 1,700
artists. Desde 280 existing different hierarchical
genres, only the G = 22 general ones on top of the
hierarchy are being kept for further analysis (p.ej.,
“Pop/General” is kept but not “Pop/Vocal Pop”).
The names of the genres plus percentages of songs
belonging to each of the genres are given in Table 1.
(Each song is allowed to belong to more than one
genre, hence the percentages in Table 1 add up to
más que 100 percent.) The genre information is
identical for all songs on an album. The numbers
of genre labels per album are given in Figure 1. Nuestro
database was set up so that every artist contributes
entre 6 y 29 albums (ver figura 2).

To study the influence of the size of the database

on results, we created random non-overlapping
splits of the entire data set: D(1/2) is two data sets
with the mean number of song excerpts = 127,199;
D(1/20) is twenty data sets with the mean number
of songs excerpts = 12,719.9; y D(1/100) is one
hundred data sets with the mean number of songs
excerpts = 2,543.98. An artist with all their albums
is always a member of a single data set.

Métodos

available preview song excerpts were obtained with
an automated Web-crawl. All meta-information
(artist name, album title, song title, genres) es

We compare two approaches based on different pa-
rameterizations of the data. Whereas mel-frequency

Flexer and Schnitzer

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

cepstral coefficients (MFCCs) are a relatively di-
rect representation of the spectral information
of a signal and therefore of the specific “sound”
or “timbre” of a song, fluctuation patterns (FPs)
are a more abstract kind of feature describing the
amplitude modulation of the loudness per fre-
quency band. It is our hypothesis that MFCCs
are more prone to pick up production and master-
ing effects of a single album as well as the spe-
cific “sound” of an artist (voice, instrumentation,
etc.).

Mel-Frequency Cepstral Coefﬁcients and Single
Gaussians (G1)

We use the following approach to music similarity
based on spectral similarity. For a given music
collection of songs, it consists of the following
steps:

1. For each song, compute MFCCs for short

overlapping frames.

2. Train a single Gaussian (G1) to model each

of the songs.

3. Compute a similarity matrix between all
songs using the symmetrized Kullback-
Leibler divergence between respective G1
modelos.

The 30-second song excerpts in MP3 format
are recomputed to 22,050-Hz mono audio sig-
nal. We divide the raw audio data into non-
overlapping frames of short duration and use
MFCCs to represent the spectrum of each frame.
MFCCs are a perceptually meaningful and spec-
trally smoothed representation of audio signals.
MFCCs are now a standard technique for com-
putation of spectral similarity in music analysis
(ver, p.ej., logan 2000). The frame size for com-
putation of MFCCs for our experiments was
46.4 mseg (1,024 muestras). We used the first 25
MFCCs for all our experiments. A G1 with full
covariance represents the MFCCs of each song
(Mandel and Ellis 2005). For two single Gaussians,
pag(X) = N (X; μ p, (cid:2) pag) and q(X) = N (X; μq, (cid:2)q), the closed
form of the Kullback-Leibler divergence is defined

como (Penny 2001):

K L N( pag(cid:3)q) = 1
2

(cid:2)

registro

det ((cid:2) pag)
det ((cid:2)q)

(cid:3)

(cid:4)
(cid:2)−1
pag

+ Tr

(cid:2)q
(cid:3)
pag (μq − μ p) − d

(cid:5)

(1)

+ (μ p − μq)

(cid:4) (cid:2)−1

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

where Tr (METRO) denotes the trace of the matrix M,
Tr (METRO) = (cid:2)i=1..nmi,i. The divergence is symmetrized
by computing:

K Lsym = K L N( pag(cid:3)q) + K L N(q(cid:3) pag)

(2)

Fluctuation Patterns and Euclidean Distance (FP)

FP (Pampalk 2001; Pampalk, Rauber, and Merkl
2002) describe the amplitude modulation of the
loudness per frequency band and are based on ideas
developed in Fruehwirt and Rauber (2001). Para
a given music collection of songs, computation
of music similarity based on FPs consists of the
following steps:

1. For each song, compute an FP.
2. Compute a similarity matrix between all

songs using the Euclidean distance of the FP
patrones.

Closely following the implementation outlined in
Pampalk (2006), an FP is computed by: (i) cutting
an MFCC spectrogram into three-second segments;
(ii) using an FFT to compute amplitude modulation
frequencies of loudness (range 0 − 10 Hz) for each
segment and frequency band; (iii) weighting the
modulation frequencies based on a model of per-
ceived fluctuation strength; y (iv) applying filters
to emphasize certain patterns and smooth the result.
The resulting FP is a 12 (frequency bands according
to twelve critical bands of the Bark scale [Zwicker
and Fastl 1999]) × 30 (modulation frequencies,
que van desde 0 a 10 Hz) matrix for each song. El
distance between two FPs i and j is computed as the
Euclidean distance:

D(FP i, FP j) =

12(cid:6)

30(cid:6)

k=1

l=1

(cid:7)
FP i
k,yo

− FP j
k,yo

(cid:8)
2

(3)

Computer Music Journal

Mesa 2. Nearest Neighbor Characteristics for
Methods G1 and FP

Método

1st AL

1st AR

AL prec

AR prec

G1
FP

27.87
2.24

35.76
26.85

13.86
0.90

8.14
1.63

1st AL = same album; 1st AR = same artist; ALprec = album
precisión; AR = artist precision.

where k indexes the frequency band and l indexes
the modulation frequency.

Resultados

In what follows, we present our results concerning
the effect of album and artist filters on album and
artist precision as well as on genre classification
actuación. This is done for the full database and
all its subsets to study the influence of the database
tamaño.

Album/Artist Precision

For the Full Database, D(ALL)

For every song in the database D(ALL), we computed
the first nearest neighbor for both methods G1 and
FP. For method G1, the first nearest neighbor is the
song with minimum Kullback-Leibler divergence
(Ecuación 2) from the query song. For method
FP, the first nearest neighbor is the song with
minimum Euclidean distance of the FP pattern
(Ecuación 3) from the query song. We then computed
the percentage of instances in which the first nearest
neighbor is from the same album (1st AL) or from
other albums by the same artist (1st AR) as the query
song (ver tabla 2).

For method G1, 27.87 percent are from the same
album and 35.76 percent from other albums by the
same artist. De término medio, hay 13.46 songs on an
album and 131.2 songs from one artist. Considered
that there are always more than 250,000 songs from
other artists, it is quite astonishing that only in 36.37
percent a song from a different artist turns up as a

Cifra 3. Percentage
(y-axis) of songs whose nth
nearest neighbor
(norte = 1 , . . . , 10 ) is from the
same album (a), or from
another album by the
same artist (b), for method
G1.

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

first nearest neighbor. For method FP, percentages
are quite a bit lower, con solo 2.24 percent from the
same album and 26.85 percent from other albums
by the same artist.

We also computed lists of the nth nearest neigh-
bors (norte = 1, . . . , 10) for every song in the database
D(ALL), for both methods, G1 and FP. Nosotros entonces
computed the percentage of instances in which
members of these lists of size n = 1, . . . , 10 son
from the same album or from other albums by
the same artist as the query song (ver figura 3
for method G1 and Figure 4 for method FP). Como
can be seen, the degradation of percentages with
growing list size for method G1 is quite graceful
(ver figura 3): Por ejemplo, the percentage of in-
stances where the five nearest neighbors are from
the same album is at 22.07 percent compared to
27.87 percent for the first nearest neighbor. El
percentage of instances where the five nearest
neighbors are from other albums by the same artist
is at 21.83 percent compared to 35.76 percent for
the first nearest neighbor. There is a similar behav-
ior for method FP at a generally lower level (ver
Cifra 4).

Next we computed the album and artist precision
at n. Album precision at n (AL prec) is the percentage
of songs from the album in a list of the n nearest
neighbors, with n being equal to the number of other

Flexer and Schnitzer

Cifra 4. Percentage
(y-axis) of songs whose nth
nearest neighbor
(norte = 1 , . . . , 10 ) is from the
same album (a), or from
another album by the
same artist (b), para
method FP.

Cifra 5. Percentage
(y-axis) of ﬁrst nearest
neighbors from the same
album (dashed line), y
from other albums by the
same artist (solid), para
different sizes of the data
colocar (x-axis, log scale), usando
method G1.

100

D(1/100)

D(1/20)

D(1/2) D(ALL)

songs in the same album as the query song. Artist
precision at n (AR prec) is the percentage of songs
from the artist in a list of the n nearest neighbors,
with n being equal to the number of other songs
from the same artist as the query song. For D(ALL)
and method G1, album precision is at 13.86 por ciento
and artist precision at 8.14 por ciento (ver tabla 2).
Precision values for method FP are very small. A
sum up, there is both an album and an artist effect
in nearest-neighbor-based music recommendation
for method G1. For this timbre-based method, el
album effect is even bigger, relatively speaking, than
the artist effect, given that, on average, the number
of songs on the same album is only a tenth of the
number of songs on other albums by the same artist.
For method FP, there is only a smaller artist effect,
and no album effect.

Inﬂuence of the Database Size

We repeated the experiments for all the subsets of
the database as described in the Data section. El
results depicted in Figures 5, 6, 7 y 8 show mean
values over 100 (D(1/100)), 20 (D(1/20)), 2 (D(1/2))
data sets or the respective single result for the full
data set D(ALL). Note that in all these figures, el
x-axis is given in log scale to better depict the large
range of values of the different sizes of data sets (de

2,543.98 for D(1/100) a 254, 398.00 for D(ALL)). El
percentage of songs whose first nearest neighbor is
from the same album decreases from 38.91 por ciento
for D(1/100) a 27.87 percent for D(ALL), for method
G1 (Cifra 5). There is a parallel decrease for the
first nearest neighbor from other albums from the
same artist for method G1 (ver figura 5). A similar
decrease at lower levels can be seen for method FP
(ver figura 6). As the data sets get larger, there clearly
seems to be an increased probability that songs from
other artists are more similar to the query song than
are songs from the same album or artist.

Album and artist precision also decrease with
increasing size of data set. For method G1, artist
precision drops from 35.99 percent for D(1/100)
a 8.14 percent for D(ALL) even falling below
album precision (ver figura 7). For method FP, artist
precision drops from 19.19 percent for D(1/100) a
1.63 percent for D(ALL) which is at the same low
level as album precision (ver figura 8). To sum up,
both first-nearest-neighbor rates and precision values
are overestimated when smaller data sets are used.

Genre Classiﬁcation

For the Full Database, D(ALL)

We also did experiments on the influence of album
and artist filters on genre classification performance.

Computer Music Journal

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Cifra 6. Percentage
(y-axis) of ﬁrst nearest
neighbors from the same
album (dashed line), y
from other albums by the
same artist (solid), para
different sizes of the data
colocar (x-axis, log scale), usando
method FP.
100

Cifra 7. Precision (y-axis)
of album (dashed line) y
artist (solid) for different
sizes of the data set
(x-axis, log scale), usando
method G1.

Cifra 8. Precision (y-axis)
of album (dashed line) y
artist (solid) for different
sizes of the data set
(x-axis, log scale), usando
method FP.

D(1/100)

D(1/20)

D(1/2) D(ALL)

Cifra 6.

100

D(1/100)

D(1/20)

D(1/2) D(ALL)

Cifra 7.

For the classifier, we used nearest-neighbor classi-
fication. For every song in the database D(ALL),
we computed the first nearest neighbor for both
methods G1 and FP. For method G1, the first nearest
neighbor is the song with minimum Kullback-
Leibler divergence (Ecuación 2) to the query song.
For method FP, the first nearest neighbor is the song
with minimum Euclidean distance of the FP pattern
(Ecuación 3) to the query song. When using an album

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

100

D(1/100)

D(1/20)

D(1/2) D(ALL)

filter (ALF), all other songs from the same album
as the query song were excluded from becoming
the first nearest neighbor. When using an artist
filter (ARF), all other songs from the same artist as
the query song were excluded from becoming the
first nearest neighbor. When using no filter (NOF),
any song was allowed to become the first nearest
neighbor. To estimate genre classification accuracy,
the genre label of a query song squer y and its first
nearest neighbor snn were compared. The accuracy is
defined as:

acc(squer y, snn) =

|gquer y ∩ gnn|
|gquer y ∪ gnn|

(4)

with gquer y being a set of all genre labels for the
query song, gnn being the analogous set for the
nearest-neighbor song, y |.| counting the number
of members in a set.

Therefore accuracy is defined as the number
of shared genre labels divided by the set size of
the union of sets gquer y and gnn. The latter is done
to penalize nearest-neighbor songs that have high
numbers of genre labels. The range of values for
accuracy is between 0 y 1. The baseline accuracy
achieved by always guessing the three most probable
genres (“Rock,” “Pop,” and “Alternative Rock,” see
Mesa 1) es 32.42 por ciento. We decided to use three
genres, rather than some other number of genres,
for this baseline accuracy, because the greatest

Flexer and Schnitzer

Cifra 9. Accuracy (y-axis)
for no ﬁlter (dotted line),
album ﬁlter (dashed line),
and artist ﬁlter (solid line),
for different sizes of the
data set (x-axis, log scale),
using method G1.

Cifra 10. Accuracy
(y-axis) for no ﬁlter (dotted
line), album ﬁlter (dashed
line), and artist ﬁlter (solid
line), for different sizes of
the data set (x-axis, registro
escala), using method FP.

100

D(1/100)

D(1/20)

D(1/2) D(ALL)

Cifra 9.

100

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

D(1/100)

D(1/20)

D(1/2) D(ALL)

Cifra 10.

Figures 9 y 10). For both methods G1 and FP, el
accuracy when using an artist filter (solid lines in
Figures 9 y 10) increases with increasing size of
data set. For G1, the increase is from 27.60 por ciento
for D(1/100) a 35.98 percent for D(ALL). For FP, él
is from 23.72 percent for D(1/100) a 28.96 por ciento
for D(ALL).

How can this contrary behavior of decreasing

accuracy for no filter and album filter versus

Mesa 3. Average Accuracies for G1 and FP without
(NOF) and with Album Filter (ALF) and Artist
Filter (ARF)

Método

G1
FP

NOF

68.98
44.35

ALF

56.07
43.03

ARF

35.98
28.96

number of songs are labeled with three genres (ver
Cifra 1). Average accuracy results for methods G1
and FP are given in Table 3. Without using any filter
(NOF), G1 clearly outperforms FP (68.98 percent vs.
44.35 por ciento). Using an album filter (ALF) strongly
degrades the performance of G1 down to 56.07
por ciento, but hardly impairs method FP. Usando un
artist filter (ARF) further degrades the performance
of G1 but also of FP. The difference between G1
and FP is now much closer (35.98 percent vs. 28.96
por ciento). Sin embargo, method G1 barely outperforms
the baseline accuracy of 32.42 percent and method
FP clearly falls below it.

Eso es, not using any filter yields very over-
optimistic accuracy results. As a matter fact, resultados
after artist filtering are very close to, or even below,
baseline accuracy. There is both an album and an
artist filter effect for G1. There is only an album
filter effect for FP. Using filters diminishes the
differences in accuracies between methods G1
and FP, since filters have a bigger impact on G1
than FP.

Inﬂuence of the Size of the Database

We repeated the experiments for all the subsets of
the database as described in the section “Data.”
The results depicted in Figures 9 y 10 show mean
accuracy values over 100 (D(1/100)), 20 (D(1/20)),
y 2 (D(1/2)) data sets or the respective single result
for the full data set D(ALL). For both methods G1 and
FP, the accuracy without using a filter (dotted lines
in Figures 9 y 10) decreases with increasing size of
data set. For G1, the decrease is from 80.65 por ciento
for D(1/100) a 68.98 percent for D(ALL). For FP, él
is from 57.19 percent for D(1/100) a 44.35 por ciento
for D(ALL). There is an almost parallel decrease in
accuracy when using album filters (dashed lines in

Computer Music Journal

increasing accuracy for artist filters be explained?
Larger data sets allow for a larger choice of songs to
become the first nearest neighbor. This larger choice
of songs can come with correct or incorrect genre
labels. If we use artist filters, this larger choice seems
to make it more probable that a song with the correct
genre label is first nearest neighbor. Otherwise we
would not see the increase in accuracy. If we use no
filter or only an album filter, the larger choice seems
to interfere with the songs from the same artist
still in the database. Songs from the larger choice
sometimes end up being first nearest neighbor
instead of a song from the same artist as the query
song. Because most songs from an artist share the
same genre labels, the larger choice in this case
diminishes the accuracy.

En otras palabras, there clearly is an influence of the
database size on accuracy performance. Small data
sets are too pessimistic when artist filters are used.
But they are overly optimistic if no filters are used,
or only album filters.

Conclusión

There are clearly both an album effect and an
artist effect in music recommendation, even in
very large databases. For the timbre-based method
G1, about one third of the first recommendations
are from the same album and about another third
from other albums by the same artist as the query
song. Considering that every artist has multiple
albums in the database and that an album contains
only about 13 songs on average, the album effect
is bigger, relatively speaking, than the artist effect.
This suggests that the direct representation of the
spectral information is more sensitive to the specific
“sound” of an album. This “sound” can be due to
instrumentation, production and mastering effects,
and other aspects that remain constant for all songs
on an individual album.

Note that we have no way to know whether an
artist is working with the same recording studio
or sound engineer for more than one album, nor
whether instrumentation, style, etc., change from
album to album. For method FP, there is only a
smaller artist effect, and no album effect. Esta sugerencia-

gests that the more abstract signal representation
of the fluctuation patterns is not sensitive to the
“sound” of individual albums. But it is still able
to model the common musical “language” of an
artist across different albums. Our experiments also
show that album and artist effects in music rec-
ommendation are overestimated when smaller data
sets are being used. Because most research on artist
filters so far concentrated on genre classification,
we also did large scale experiments on classification
exactitud. We corroborated earlier results that not
using any filter yields very over-optimistic accuracy
resultados. Using artist filters even reduces results
close to or below baseline accuracy. As reported
antes, using artist filters also diminishes the dif-
ferences in accuracies between methods that are
affected distinctly by filtering. Además, allá
clearly is an influence of database size on accuracy
actuación.

As with all large-scale performance studies, allá
remains the question as to how representative and
universally valid our results are. We are convinced
that our database is representative of music that is
generally listened to and available in the Western
hemisferio, since it is a large and random subset of
acerca de 5 million songs from a popular Web store. Como
to the methods employed, we chose one method that
closely models the audio signal and one that extracts
information on a somewhat higher level. It is our
guess that other methods’ performance will be close
to either of our methods, depending on their level
of closeness to the analyzed audio signal. Sistemas
using a Gaussian representation of spectral features
together with the Kullback-Leibler divergence (como
our method G1) regularly rank in the top places of
the yearly Music Information Retrieval Evaluation
Exchange (MIREX, www.music-ir.org/mirexwiki/)
(Downie 2008), which is a community-based frame-
work for the formal evaluation of music information
retrieval systems and algorithms. This shows that
at least one of our chosen music recommendation
methods belongs to the most successful approaches
in music similarity. The choice of our methods was
also influenced by considerations of computability.
Después de todo, 250,000 song excerpts are a lot of data to
analizar, and both our methods can be implemented
very efficiently. Using nearest-neighbor methods for

Flexer and Schnitzer

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

music recommendation seemed to be the obvious
choice.

With audio-based music recommendation ma-
turing to the scale of the Web, our work provides
important insight into the behavior of music simi-
larity for very large databases. Even with hundreds of
thousands of songs, album and artist effects remain
a problem.

Expresiones de gratitud

Parts of this work were funded by the Austrian
Fundación Nacional de Ciencia (FWF) (Projects P21247
and L511-N15) and by the Austrian Research
Promotion Agency (FFG) (Bridge-project 815474).

Referencias

Downie, j. 2008. “The Music Information Retrieval
Evaluation Exchange (2005–2007): A Window into
Music Information Retrieval Research.” Acoustical
Science and Technology 29(4):247–255.

Flexer, A. 2006. “Statistical Evaluation of Music

Information Retrieval Experiments.” Journal of New
Music Research 35(2):113–120.

Flexer, A. 2007. “A Closer Look on Artist Filters for
Musical Genre Classification.” Proceedings of the
Eighth International Conference on Music Information
Retrieval (ISMIR ’07). Viena: Austrian Computer
Sociedad, páginas. 341–344.

Fruehwirt, METRO., y un. Rauber. 2001. “Self-Organizing

Maps for Content-Based Music Clustering.” Proceed-
ings of the Twelth Italian Workshop on Neural Nets.
Londres: Saltador, páginas. 228–233.

logan, B. 2000. “Mel Frequency Cepstral Coefficients

for Music Modeling.” Proceedings of the International
Symposium on Music Information Retrieval (ISMIR
’00). Available at http://ismir2000.ismir.net/. Last
accedido 26 Abril 2010.

Mandel, METRO., y D. Ellis. 2005. “Song-Level Fea-
tures and Support Vector Machines for Music
Classification.” Proceedings of the Sixth Interna-
tional Conference on Music Information Retrieval
(ISMIR ’05). Londres: University of London, páginas. 594–
599.

Pampalk, mi. 2001. “Islands of Music: Análisis, Organiza-
ción, and Visualization of Music Archives.” MSc thesis,
Technical University of Vienna.

Pampalk, mi. 2006. “Computational Models of Music

Similarity and Their Application to Music Information
Retrieval.” Doctoral thesis, Vienna University of
Tecnología.

Pampalk, MI., A. Flexer, y G. Widmer. 2005. “Improve-
ments of Audio-Based Music Similarity and Genre
Classification.” Proceedings of the Sixth International
Conference on Music Information Retrieval (ISMIR
’05). Londres: University of London, páginas. 628–
633.

Pampalk, MI., A. Rauber, y D. Merkl. 2002. “Content-
Based Organization and Visualization of Music
Archives.” Proceedings of the Tenth ACM Interna-
tional Conference on Multimedia. Nueva York: ACM,
páginas. 570–579.

Penny, W.. D. 2001. “Kullback-Leibler Divergences of
Normal, Gamma, Dirichlet and Wishart Densities.”
Technical report, Wellcome Department of Cognitive
Neurología, University College London.

Zwicker, MI., and H. Fastl. 1999. Psychoacoustics, Facts
and Models. Springer Series of Information Sciences,
volumen 22. Berlina: Saltador, 2nd edition.

Computer Music Journal

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh
metro

j
/

a
r
t
i
C
mi
–
pag
d

F
/

3
4
3
2
0
1
8
5
5
5
8
4
/
C
oh
metro
_
a
_
0
0
0
0
4
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
Descargar PDF