Musical Instrument Classification Using Neural Networks

MUSTAFA SARIMOLLAOGLU1 and COSKUN BAYRAK2
1Dept. of Applied Science, 2 Dept. of Computer Science
University of Arkansas at Little Rock
2801 S. University Avenue Little Rock, Arkansas 72204
USA


Abstract: - In this paper, a system for automatic classification of musical instrument sounds is introduced. As
features mel-frequency cepstral coefficients and as classifiers probabilistic neural networks are used. The
experimental dataset included 4548 solo tones from 19 instruments of MIS database (The University of Iowa
Musical Instrument Samples). Experiments for different system structures (hierarchical and direct
classification) were carried out and compared. The best performance in direct classification was 92% for
individual instruments and 97% for families; and 89% for individual instruments when hierarchical approach
is used.

Key-Words: - Musical instrument classification, probabilistic neural networks, PNN

1 Introduction
Music Information Retrieval (MIR) has gained
increasing research attention over the recent years.
Apart from their academic merits, robust MIR
systems will have important commercial and social
implications. They will add significant value to the
existing music libraries by making them easily
accessible; enabling automatic classification,
organization, indexing, and searching. Musical
instrument classification, where the idea is to
recognize the instruments playing in a musical
sound, is one of the signal analysis problems in
MIR.
Musical instrument recognition is a difficult task
and is far from being solved and applicable to real-
world musical signals [1]. The problem is rather
easy for monophonic sounds compared to
polyphonic ones, where multiple instruments played
together. Assuming a preliminary source separation
has been performed, classification research has been
concentrated on monophonic sounds.
The recognition of audio signals consists of two
basic steps; defining and extracting the features that
distinguish the sources, and design of a system
(classifier) to recognize the sources using those
features. Many features (cepstral, spectral, temporal)
are introduced in the literature and a comparison of
them with regard to recognition performance can be
found in [2]. K-Nearest Neighbors (k-NN), Hidden
Markov Models, Gaussian Mixture Models (GMM),
Naive Bayesian classifiers, Support Vector
Machines, and Artificial Neural Networks (ANN)
are some of the techniques used for instrument
classification.
Eronen and Klapuri’s system used a wide set of
features and tested on 30 instruments [3]. They
utilized a hierarchical framework for classification
and used Gaussian or k-NN classifier at each node
but direct classification performed better. Best
recognition accuracy was 94% for instrument family
and 80% for individual instruments. Krishna and
Sreenivas proposed line spectral frequencies as
features and obtained 95% and 90% accuracy for
family and instruments respectively [4]. They
classified 14 instruments using GMMs and K-NN
classifiers. Bolat compared the performances of
three statistical neural networks, on four reed
instruments using linear prediction coefficients as
features [5]. PNN achieved highest accuracy of 93%
compared to GRNN (90%) and RBF (47%)
networks.
In this paper we focus on the performance of
probabilistic neural networks on classification of a
large number of instruments, using only the cepstral
features.


2 Feature Extraction
Mel-Frequency Cepstral Coefficients (MFCCs) are
the features used in our system to model the tones.
MFCCs have been proven to be useful in a broad
range of classification applications, such as speech
classification [6], speaker identification [7], musical
genre classification [8], etc. In [2], a large set of
features are compared in terms of recognition
performance in an instrument recognition system.
Among the others, MFCCs, their standard

Proceedings of the 5th WSEAS International Conference on Signal Processing, Istanbul, Turkey, May 27-29, 2006 (pp151-154)

Figure 1. Computation of MFCCs

deviations, and deltas seemed to be the most
successful ones [2].
MFCCs provide a compact representation of the
spectral envelope and are extracted as given in
Figure 1. Mel-scaling emphasizes the perceptually
meaningful frequencies by mapping the spectrum
coefficients into a non-linear manner. Discrete
cosine transform (DCT) is used to reduce the
dimension of power spectrum.
In this work, the input signal was processed in
256 point frames overlapped by 50 %. First 12
MFCCs (excluding 0th) were calculated. Resulting
12-dimensional vectors were used as the base
feature vectors. Keeping the same file structure as in
MIS, we had several feature files for each
instrument.
Prior to the classification stage, vector
quantization was applied to further reduce the
amount of data and complexity. LBG algorithm,
which is an efficient variant of k-means algorithm,
was used in quantization [9].
Three sets of quantized sample vectors were
produced to be used in appropriate experiments. In
the first step, feature vectors in every feature file
were individually clustered to 100 vectors (Set 1). In
the second step, for each instrument, 70% of the
feature files from Set 1 were selected randomly and
they were merged and clustered into 300 vectors to
represent each instrument (Set 2). In the last step,
samples in Set 2 were combined within the
instrument families and clustered into 300 vectors
(Set 3). Cluster size was chosen to be 300 as a trade-
off between computational complexity and a good
representation of the class. Remaining 30% of the
Set 1 samples were left for testing purpose while Set
2 and 3 were used for training. Hence, testing and
training samples were completely different.


3 Classifier
Artificial Neural Networks (ANNs) are processing
structures that consist of interconnected neurons.
Connectivity pattern and weights between neurons
represents the mappings between input and output
vectors. Some ANNs have the ability of
approximating any function, but in general it takes
very long to train the network and adjust the
parameters [1].
Probabilistic Neural Network (PNN) is a type of
statistical neural networks. PNN learns to
approximate the probability density function of the
training examples. The only parameter that needs to
be selected for training is the spread, which is the
deviation of the Gaussian functions. Spread is
chosen experimentally to find the best results. For
more information about PNN, please refer to [7].
PNNs’ advantages over the other methods are
flexibility and the straightforward design. Training
time is much faster than the other types of ANNs.
They enable incremental training, where new
training examples can be incorporated without
difficulties. And they are robust to noise. However,
the major disadvantage of the PNN is that it is
slower to operate, because it performs more
computations then other ANN models [10].
In our work, PNNs were used for classification.
In the non-hierarchic case, one network was used.
The hierarchic scheme composed of five networks in
two stages; one for family classification and four for
instrument classification within the family.

Decision Mechanism
PNN provides best matching instrument for each
input vector. However, in order to increase the
accuracy, multiple input vectors from each sample
are needed. At the decision level, a predefined
number of outputs from the PNN are buffered and
summed together. This way, each class in the system
gets a score on the period of sample applied to the
system. Then the class with the highest score is
chosen to be the source of the sample [7].


4 Experiments and Results
Sample database used in experiments consist of
4548 tones from 19 instruments of MIS database, as
detailed in Table 1. All samples are in mono, 16 bit
and 44.1 kHz. These samples include several
different articulation styles; all strings include
pizzicato and some instruments include vibrato. All
instruments have samples of three dynamic levels
(ff,mf,pp).
Two systems were implemented; one with direct
classification and one with hierarchical
classification. Direct classification system has one
PNN, having 19 classes, which are as many as
instruments. It is trained with features from Set 2
(detailed in section 2). The hierarchical system on
the other hand, has a structure consisting of two
Proceedings of the 5th WSEAS International Conference on Signal Processing, Istanbul, Turkey, May 27-29, 2006 (pp151-154)