~ICMC 2015 - Sept. 25 - Oct. 1, 2015 - CEMI, University of North Texas
12 12 12 12
6 6 6 6
0 0 0 0
6 -6 -6 -6
12 -12 -12 -12
Classical Disco Jazz Metal
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
12 12 12
6 6 6
0 0 0
-6 -6 -6
-12 -12 -12
Pop
Reggae
Rock
Figure 1. Equalization curves as proposed by the VLC media player for the musical genres Classic, Disco, Jazz, MetalPop, Reggae and Rock. The EQ has 10
bands, centered on 60, 170, 310, 600, 1000, 3000, 6000, 12000, 14000 and 16000 Hz. Pictures shows the gains on dB in a range from -12 to 12 dB.
theories and technologies that employ fuzzy sets [8]. In gen- We used the GTZAN Genre Collection2, a well known database
eral, when fuzzy logic is applied to computers, it allows them in the Music Information Retrieval community. It contains
to emulate the human reasoning process, quantify imprecise one hundred audio tracks of thirty seconds long of each of the
information and make decisions based on vague and incom- following genres: Blues, Classical, Country, Disco, Hiphop,
plete data [5]. Jazz, Metal, Pop, Reggae and Rock. Ninety tracks per genre
Fuzzy systems are powerful and work in a way that resem- were select to be analyzed and the other ten were reserved for
bles some characteristics of human behavior. Parallel compu- testing.
tation of fuzzy rules reduces drastically the computation time We checked the usability of these descriptors for genre clascompared to a traditional mathematical approach. Fuzzy sys- sification. We conducted analysis of variance (ANOVA) tests
tems allow approximation of highly nonlinear systems with for every descriptor. This statistical test determines signifihigh accuracy. It is not necessary to know any mathemati- cant differences in groups of data [18]. It estimates the varical model in advance to approximate any system [5]. Fuzzy ance on a normal distribution for each group and provides a
logic allows us to build systems using common sense, and parameter p that represents the probability that two or more
the fuzzy rules can be discussed, tuned, and detuned easily. sets belong to the same group. A value less than 0.01 deterThese facts makes fuzzy logic a very appropriate tool to study mines that groups are independent sets. For each genre, 90
and model non-linear dynamical systems and to handle com- songs were selected and their audio descriptors were complexity.Fuzzy logic systems have been widely used in engi- puted. The highest p-value obtained was 0,000239. This
neering and control applications [4] [5], and also in several means that the eighteen chosen descriptors have a small probartistic and musical applications such as sound synthesis [9], ability of belonging to the same group, and we than deduct
polyphonic note extraction [10], digital restoration [11], mu- that these genres could be indeed characterized by the chosen
sical decisions [12], and audio-visual composition [13]. parameters.
4. AUDIO EQUALIZATION BASED ON FUZZY
3. GENRE DESCRIPTORS LOGIC
Some authors have proposed the usage of audio descriptors We propose an equalization system as the one described in
to identify different musical genres [14-16]. [16] shows tim- figure 2. We start by computing the eighteen audio descripbre descriptors like a useful way to classify genres, so we use tors of section 3 in a window of 30 seconds, using the MIR
these descriptors to detect when a song belongs to one genre toolbox [17]. As previously discussed, these descriptors are
or another. Based on this knowledge, we decided to use the supposed to represent specific aspects of different musical
MIR toolbox, a very complete musical information retrieval genres and styles. We then fuzzyfy these descriptors, and
analysis tool for MATLAB 1 The descriptors we used were use them as inputs for a fuzzy logic inference engine. Uszerocross, centroid, low energy, rolloff, flux and MFCC coef- ing several decision rules derived from the descriptor analyficients [17]. ses (described in section 4.3), the fuzzy system determines
2available from http://marsyas.info/downloads/
1 http: //www.mathworks.com datasets.html
- 135 -
0
