Music Instrument Estimation in Polyphonic

Sound Based on Short-Term Spectrum Match

Wenxin Jiang1 , Alicja Wieczorkowska2 , and Zbigniew W. Raś1 ,2

1
University of North Carolina,
Department of Computer Science, Charlotte, NC 28223, USA
2
Polish-Japanese Institute of Information Technology,
Koszykowa 86, 02-008 Warsaw, Poland
wjiang3@uncc.edu, alicja@pjwstk.edu.pl, ras@uncc.edu

Summary. Recognition and separation of sounds played by various instruments is

very useful in labeling audio files with semantic information. This is a non-trivial
task requiring sound analysis, but the results can aid automatic indexing and brows-
ing music data when searching for melodies played by user specified instruments. In
this paper, we describe all stages of this process, including sound parameterization,
instrument identification, and also separation of layered sounds. Parameterization
in our case represents power amplitude spectrum, but we also perform compar-
ative experiments with parameterization based mainly on spectrum related sound
attributes, including MFCC, parameters describing the shape of the power spectrum
of the sound waveform, and also time domain related parameters. Various classifi-
cation algorithms have been applied, including k-nearest neighbor (KNN) yielding
good results. The experiments on polyphonic (polytimbral) recordings and results
discussed in this paper allow us to draw conclusions regarding the directions of fur-
ther experiments on this subject, which can be of interest for any user of music
audio data sets.

1 Introduction

Recently, a number of acoustical features for the construction of a compu-

tational model for music timbre estimation have been investigated in Music
Information Retrieval (MIR) area. Timbre is a quality of sound that distin-
guishes one music instrument from another, since there are a wide variety of
instrument families and individual categories. It is rather a subjective quality,
defined by ANSI as the attribute of auditory sensation, in terms of which a
listener can judge that two sounds, similarly presented and having the same
loudness and pitch, are different [1], [2]. Such definition is clearly subjective
and not of much use for automatic sound timbre classification, although the
footnote to the definition gives hints towards physical timbre description, sta-
ting that the timbre depends primarily upon the spectrum of the stimulus,
2 Wenxin Jiang, Alicja Wieczorkowska, and Zbigniew W. Raś,

but also upon the waveform, the sound pressure, the frequency location of the
spectrum, and the temporal characteristics of the stimulus [2], [5]. Still, mu-
sical sounds must be very carefully parameterized to allow automatic timbre
recognition.
So far, there is no standard parameterization used as a classification basis.
The sound descriptors applied are based on various methods of analysis in
time domain, spectrum domain, time-frequency domain and cepstrum, with
Discrete Fourier Transform (DFT) for spectral analysis being most common,
e.g. Fast Fourier Transform (FFT), and so on. Also, wavelet analysis gains
increasing interest for sound and especially for musical sound analysis and
representation.
Researchers explored different statistical summations to describe signa-
tures of music instruments based on vectors or matrices in features, such as
Tristimulus parameters, brightness, irregularity of the spectrum etc. [6], [14],
[21]. Flattening these features for traditional classifiers increases the number
of features. In [16] authors used a new set of features jointly with other pop-
ular features used in music instrument identification. They built a database
of music instrument sounds for training a number of classifiers. These classi-
fiers are used by MIRAI system to identify music instruments in polyphonic
sounds.
MIRAI is designed as a web-based storage and retrieval system which
can automatically index musical input (of polyphonic, polytimbral type),
transforming it into a database, and answer queries requesting specific musi-
cal pieces, see http://www.mir.uncc.edu/. When MIRAI receives a musical
waveform, it divides this waveform into segments of equal size and then the
classifiers incorporated into the system identify the most dominating musical
instruments and emotions associated with that segment. A database of mu-
sical instrument sounds describing about 4,000 sound objects by more than
1,100 features is associated with MIRAI. Each sound object is represented as
a temporal sequence of approximately 150-300 tuples which gives a temporal
database of more than 1,000,000 tuples, each one represented as a vector of
about 1,100 features. This database is mainly used to train classifiers for auto-
matic indexing of musical instrument sounds. It is semantically reach enough
(in terms of successful sound separation and recognition) so the constructed
classifiers have a high level of accuracy in recognizing the dominating mu-
sical instrument and/or its type when music is polyphonic. Unfortunately,
the loss of information on non-dominant instruments by the sound separation
algorithm, due to the overlap of sound features, may significantly lower the
recognition confidence of the remaining instruments in a polyphonic sound.
This paper shows that by identifying a weighted set of dominating instruments
in a sequence of overlapping frames and using a special voting strategy, we can
improve the overall confidence of the indexing strategy for polyphonic music,
and the same improve the precision and recall of MIRAI retrieval engine.
 Title Suppressed Due to Excessive Length 3

2 Sound Parameterization for Automatic Classification

Purposes
Sound wave can be described as a function representing amplitude changes in
time. For digitally recorded sound, this function is quantized in time and in
amplitude. Sampling rate describes how many values are recorded per time
unit, and binary resolution in multi-bit recording describes how many bits are
used to represent quantized amplitude axis, for each channel. Standard setting
for CD for sampling rate is 44,100 samples per second, i.e. 44.1 kHz, with 16-
bit resolution for amplitude, i.e. with 216 quantization levels. Obviously, such
data are not well suited to perform automatic classification, so usually sound
parameterization is performed before further experiments on audio databases.
There are numerous ways the sound can be parameterized. Audio data can
be parameterized in time domain, in frequency domain, and time-frequency
features can also be extracted. Parameterization can be based on Fourier
analysis, particularly on amplitude spectrum, on cepstral analysis, also on
wavelet analysis, and so on; various features can be extracted to describe the
results of these analyzes [3], [4], [7], [8], [9], [11], [13], [19], [22].
In our research, we decided to base mainly on direct observation of sound
spectrum (see Section 4), but we also performed experiments based on the
following sound features [8], [22]:
• AudioSpectrumBasis - MPEG-7 descriptor, representing low-dimensional
projection of a high-dimensional spectral space, to aid compactness and
recognition [8]. AudioSpectrumBasis is a matrix derived from the SVD
(singular value decomposition) of a power spectrum in normalized dB
scale, i.e. in log scale with maximal value defining 0 dB. In our research,
frequency axis for AudioSpectrumBasis was divided into 32 bands, with
1/4-octave resolution for 8 octaves; octave distance means doubling the
fundamental frequency, i.e. the pitch of the sound.
• AudioSpectrumP rojection - projection of AudioSpectrumBasis [8],
• AudioSpectrumF latness - MPEG-7 parameter, calculated in our research
for spectrum divided into 32 frequency bands, i.e. with 1/4-octave resolu-
tion for 8 octaves, and the length of this 32-element vector is added as 0th
element of this 33-dimensional feature; if there is a high deviation from a
flat spectral shape for a given band, it may signal the presence of tonal
components [8],
• M F CC = {mf ccn : 1 ≤ n ≤ 13} - cepstral coefficients in mel scale; feature
originating from speech processing, but also used for music analysis [12],
[17]. 13 coefficients were used (the 0th one and the next 12), for 24 mel
frequency scale hearing filters, using Julius software [10],
• HamonicP eaks = {HamoP kn : 1 ≤ n ≤ 28} - sequence of the first 28
local peaks of harmonics (in normalized dB scale) for a given frame
• T emporalCentroid - time instant where the energy of the sound is focused,
calculated as energy weighted mean of the sound duration,
4 Wenxin Jiang, Alicja Wieczorkowska, and Zbigniew W. Raś,

• LogSpecCentroid - AudioSpectrumCentroid from MPEG-7 standard [8];

this parameter represents the gravity center of a log-frequency power spec-
trum,
• LogSpecSpread - AudioSpectrumSpread descriptor from MPEG-7 [8];
calculated as RMS (Root Mean Square) value of the deviation of the power
spectrum in log frequency scale with respect to the gravity center in a
frame,
• Energy - energy of spectrum, averaged through all frames of the sound,
• ZeroCrossings - zero-crossing rate, i.e. number of sign changes of the
wave form in a frame, averaged through all frames of the sound,
• SpecCentroid - calculated as HarmonicSpectralCentroid from MPEG-
7, representing power-weighted average of the frequency of the bins in the
linear power spectrum, and averaged over all the frames of the steady state
of the sound,
• SpecSpread - calculated as HarmonicSpectralSpread from MPEG-7, de-
scribing the amplitude-weighted standard deviation of the harmonic peaks
of the spectrum, normalized by the instantaneous HarmonicSpectralCentroid
and averaged over all the frames of the steady state of the sound,
• RollOf f - averaged (over all frames) frequency below which an experimen-
tally chosen percentage of the accumulated magnitudes of the spectrum is
concentrated,
• F lux - difference between the magnitude of the amplitude spectrum points
in a given and successive frame, averaged through the entire sound,
• LogAttackT ime - decimal logarithm of the sound duration from the time
instant when the signal starts, to the time when it reaches its maximum
value, or when it reaches its sustained part, whichever comes first.

3 Polyphonic Sound Estimation Based on Sound

Separation and Feature Extraction
The traditional way of pattern recognition in music sound is to extract fea-
tures from raw signals in digital form, usually recorded as a sequence of integer
samples representing quantized values of amplitude of a sound wave in conse-
quent time instants. By feature extraction, the acoustic characteristics such as
pitch and timbre are described by smaller and more structured dataset which
is further fed to traditional classifiers to perform estimation.
In case of polyphonic sounds, sound separation can be applied to extract
the signal which is identified as one specific instrument at timbre estimation
process. Then timbre estimation can be applied again on the residue of the
signal to get other timbre information. Fig. 1 shows the process of music
instrument recognition system based on feature extraction and sound separa-
tion.
However, there are two main problems in this method. First of all, overlap-
ping of the features make it difficult to perform timbre estimation and sound
 Title Suppressed Due to Excessive Length 5

Polyphonic
Sound

Get pitch

FFT Feature Pitch Classifier

extraction Estimation

Get spectrum
Timbre
Estimation

Sound Get Instrument

separation
New spectrum

Power Spectrum

Fig. 1. Flow chart of music instrument recognition with sound separation

separation. Secondly, during the classification process, only one instrument is

picked up from all candidates, which makes the estimation inaccurate.

3.1 Overlapping of the Features Both in Temporal and Spectral

Space

Feature based datasets are easier and more efficient to work with classi-
fiers, however, there is usually information loss during the feature extraction
process. Feature is the abstract or compressed representation of waveform or
spectrum, such as harmonic peaks, MFCC (Mel Frequency Cepstral Coeffi-
cients), zero-crossing rate, and so on. In the case of monophonic music sound
estimation tasks with only singular non-layered sounds, the features can be
easily extracted and identified. However, this is not the case in polyphonic,
polytimbral sound. It is difficult or even often impossible to extract distinct
clear features representing single instrument from polyphonic sound, because
the overlapping of the signals and their spectra, especially when instruments
have the similar patterns in their features space.

3.2 Classification with Single Instrument Estimation for Each

Frame

The traditional classification process usually gives single answer, representing

one class; in our case, it would be the name of instrument playing in an an-
alyzed sample. In such a case, when only the best answer, i.e. the name of
6 Wenxin Jiang, Alicja Wieczorkowska, and Zbigniew W. Raś,

one (the only one or dominating) instrument playing for each frame of music
sound is given, then information about other possibly contributing instru-
ments is lost.
In fact, it is common for the polyphonic music sound to have multiple
instruments playing simultaneously, which means that in each frame, there
are representations of multiple timbres existing in the signal. Providing one
only candidate leads to obtaining predominant timbre while ignoring other
timbre information. And also, there could be no dominating timbre in each
frame, when all instruments play equally loud. This means that classifier has
to randomly choose one of the equally possible candidates. In order to find
solution to this problem, authors introduce the Top-N winner strategy which
gives multiple candidates for each evaluated frame.

4 Pattern Detection Directly from Power Spectrum

The fact that discriminating one instrument from another depends on more
details from raw signals leads to another way of pattern recognition: directly
detecting distinct patterns of instruments based on lower representation of
signal, such as power spectrum. Fig. 2 shows two different ways of pattern
recognition.

Fig. 2. Two different methods of pattern recognition

 Title Suppressed Due to Excessive Length 7

Since spectrum is very useful for timbre representation purposes, we pro-

pose the new strategy of instrument estimation based on short term power
spectrum match.

4.1 Sub-Pattern of Single Instrument in the Mixture Sound

Segment

Figure 3 shows the power spectrum of trumpet, piano and the mixture of
those two instruments. As we can see, the spectrum of mixture preserves part
of the pattern of each single instrument.

Fig. 3. Power spectrum of trumpet, piano and their mixture; frequency axis is in
linear scale, whereas amplitude axis is in log [dB] scale

The same similarity of properties of the spectra is also observed e.g. for
flute, trombone and their mixture, as Figure 4 shows.
In order to index the polyphonic sound, we need to detect the instrument
information in each small slice of music sound. Such detection is rather not
feasible directly in time domain. Therefore, in our experiments, we have ob-
served the short term spectrum space. This has been calculated via short time
8 Wenxin Jiang, Alicja Wieczorkowska, and Zbigniew W. Raś,

Fig. 4. Power spectrum of flute, trombone and their mixture

Fourier transform (STFT). Figure 5 shows another example of the spectrum

slice for flute and trombone and their mixture sound. Each slice is 0.04 seconds
long.
As Figure 5 shows, the power spectrum patterns of single flute and single
trombone can still be identified in mixture spectrum without blurring with
each other (as marked in the figure). Therefore, we do get the clear picture
of distinct pattern of each single instrument when we observe each spectrum
slice of the polyphonic sound wave.

4.2 Classification Based on Power Spectrum Pattern Match

In order to represent accurately the short term spectrum with high resolution
in frequency axis, allowing more precise pattern matching, long analyzing
frame with 8192 numeric samples was chosen. Fourier transform performed
on these frames describes frequency space for each slice (or frame). Instead of
parameterizing the spectrum (or time domain) and extracting a few dozens
 Title Suppressed Due to Excessive Length 9

Fig. 5. Sub-patterns of single instruments in the mixture sound slice for flute,
trombone, and their mix

of features to represent sound, we decided to work directly on the power

amplitude spectrum values (points). When a new sound is analyzed with a
goal to find what instrument or instruments contributed to create this sound,
even though their spectra overlap, we can still try to find the closest vectors
from the training data set of singular sounds and discover which instrument
sounds they represent.
The traditional classification models such as decision trees, Naive Bayesian
classifiers, and neural networks do not perform well in this case. It is because
there are too many attributes (8192 numeric attributes) for those classifiers
10 Wenxin Jiang, Alicja Wieczorkowska, and Zbigniew W. Raś,

to yield good classification models, and also any classification model itself
stands for some sort of abstraction, which is in conflict with any information
preserving strategy. However, one of the most fundamental and simple clas-
sification methods, K Nearest Neighbor algorithm, needs no prior knowledge
about the distribution of the data and it seems to be an appropriate classifier
for numeric spectrum vectors.

5 Top-N winners from classification model

As opposed to the traditional pattern matching or classification process, which
uses classification model to choose the one with the highest confidence as the
estimation result, we do not get the ”best answer” for an evaluated frame.
Instead, we choose the multiple candidates from classification model according
to the confidence measure. As it was already discussed, we know that the
classifier during the classification process compares pattern from each frame
with standard instrument patterns in a training database. Since there are
several different sub-patterns existing in the spectrum, classifier will assign
the confidence to each recognized pattern. This way, we may identify which
N matches have the highest confidence levels. They are our top N winners.
Thus, at each frame we get n instruments Ii with the confidence level Ci
and save them to the candidates pool for the voting process. After evaluating
all the frames, we get weights for all the candidates from the candidates pool
by adding up their confidences, and the final voting proceeds according to
the weights Wj of each instrument. The following is the pseudo-code for the
Top-N winners procedure:
For each frame from the sound
Get power spectrum by STFT
For each candidate Xi from top-N winners of classifiers
If Xi exists in candidates pool then
Confidence[x] += Ci
Else
Add Xi into candidates pool
Confidence[x] = Ci
End If
End For
End For
Select Top m candidates from candidates pool
Some noise coming from errors occurred during the single frame estimation
process could be minimized in terms of the whole music context. By keeping
the original acoustical information of the music sound, we are getting much
higher recognition rate for multiple instruments in polyphonic sound.
Here are the steps of pattern matching process:
1. Use STFT and Hamming window to extract power spectrum for each 0.04s
frame for all the standard single instrument sounds.
 Title Suppressed Due to Excessive Length 11

2. Save these spectra in a training database; since there is overlapping of 2/3

of frame length for Hamming window, the number of items in the dataset
actually almost triples for each sound.
3. During the estimation process, use KNN to do the vector distance measure
(8192 points) and decide which frame in the training dataset is the most
similar to the unknown sound frame; when we give multiple matches, the
multiple instrument candidates are saved for the overall weights calcula-
tion.
Fig. 6 shows the new music instrument recognition system which has been
developed with the strategy of Top-N winners based on short-term spectrum
matching.

Polyphonic
Sound

Training data

Get frame
Timbre
Estimation
Classifier
Get Instrument
FFT Power Spectrum
Candidates

Finish all the Frames

Get Final winners Voting process estimation

Fig. 6. Flow chart of music instrument recognition system with new strategy

6 Experiment of Top-N-winners strategy Based on Short

Term Spectrum Matching
To simplify the problem, we only performed tests on the middle C instrument
sounds, i.e. for pitch equal to C4 in MIDI notation, of frequency 261.6 Hz (for
A4 tuned to 440 Hz). The training subset including 3323 objects has been
selected from the entire training database. Each object is represented by the
frame-wise (0.04 seconds) power spectrum extracted by short time Fourier
transform from the following 26 single instrument sounds:
12 Wenxin Jiang, Alicja Wieczorkowska, and Zbigniew W. Raś,

Electric Guitar, Bassoon, Oboe, B-flat Clarinet, Marimba, C Trumpet, E-

flat Clarinet, Tenor Trombone, French Horn, Flute, Viola, Violin, English
Horn, Vibraphone, Accordion, Electric Bass, Cello, Tenor Saxophone, B-Flat
Trumpet, Bass Flute, Double Bass, Alto Flute, Piano, Bach Trumpet, Tuba,
and Bass Clarinet.
To compare the results with the traditional feature based classification
strategy, we have also extracted the following 5 groups of both temporal and
spectral features (calculated for spectrum divided into 33 frequency bands),
mainly originating from the MPEG-7 standard [8], [20], [22], and used decision
tree classifier to perform the timbre estimation:
Group1: BandsCoef f icient = {bandsCoefn : 1 ≤ n ≤ 33} - coefficients
for 33 AudioSpectrumF latness bands.
Group2: P rojections = {prjn : 1 ≤ n ≤ 33} - AudioSpectrumP rojection
from MPEG-7,
Group3: M F CC = {mf ccn : 1 ≤ n ≤ 13}
Group4: HamonicP eaks = {HamoP kn : 1 ≤ n ≤ 28}
Group5: Other Features:
• T emporalCentroid,
• LogSpecCentroid,
• LogSpecSpread,
• Energy,
• ZeroCrossings,
• SpecCentroid,
• SpecSpread,
• RollOf f ,
• F lux,
• bandsCoef Sum - AudioSpectrumF latness bands coefficients sum,
• prjmin, prjmax, prjsum, prjdis, prjstd - minimum, maximum, sum, dis-
tance, and standard deviation of AudioSpectrumP rojection calculated for
AudioSpectrumBasis. Distance represents a dissimilarity measure: dis-
tance for a matrix is calculated as sum of absolute values of differences
between elements of each row and column. Distance for a vector is calcu-
lated as the sum of dissimilarity (absolute difference of values) of every
pair of coordinates in the vector,
• LogAttackT ime.
52 polyphonic audio files have been mixed (using Sound Forge sound editor
[18]) from 2 of those 26 instruments sound. These mixture audio files have been
used as test files.
The system uses MS SQLSERVER2005 database to store training dataset
and K nearest neighbor algorithm as the classifier. When the polyphonic sound
is submitted to system, it provides several estimations as the final candidate
answers. In our experiment, we gave 4 estimations for each submitted audio
file.
 Title Suppressed Due to Excessive Length 13

The performance of our algorithm was measured using recognition rate R,

calculated as
R = P/A
where P is the positive response, i.e. the number of the correct estimations,
and A is the actual number of instruments existing in the polyphonic sound.
For comparison purpose, five experiments were performed independently.
We applied feature-based sound separation strategy and we used a decision
tree type classifier in our first two experiments. In experiment 1, only one can-
didate was chosen by a classifier for each frame. In the first step of experiment
2, top n candidates (with n = 2) were chosen by a classifier for each frame. In
its second step, for each candidate, the confidences over all the frames were
added to get the overall score used to identify the final n winners.
In the remaining three experiments, we applied a new strategy of spectrum
match based on KNN classifier. In experiment 3, we used KNN (k = 1) to
choose the top 2 candidates as the winners for each frame. In experiment 4,
we increased k from 1 to 5. In experiment 5, we ruled out the percussion
instrument objects from the testing audio files, since they have less clear
patterns in the spectrum envelope.

Table 1. Recognition rate of music instrument estimation based on various strate-

gies
experiment Recognition
description
# Rate
1 Feature-based and separation + Decision Tree (n=1) 36.49%
2 Feature-based and separation + Decision Tree (n=2) 48.65%
3 Spectrum Match + KNN(k=1;n=2) 79.41%
4 Spectrum Match + KNN(k=5;n=2)_ 82.43%
Spectrum Match + KNN(k=5;n=2) without percussion
5 instrument_
87.1%

From the results shown in Table 1, we get the following conclusions:

1. Using the multiple candidates for each frame yields better results than
using single winner strategy.
2. Spectrum-based KNN classification improves the recognition rate of poly-
phonic sounds significantly.
3. Some percussion instrument (such as vibraphone, marimba) are not suit-
able for spectrum-based classification, but most instruments generating
harmonic sounds work well with this new strategy.

7 Conclusion
We have provided a new solution to an important problem of instrument
identification in polyphonic music: The loss of information on non-dominant
14 Wenxin Jiang, Alicja Wieczorkowska, and Zbigniew W. Raś,

instruments during the sound separation process due to the overlapping of

sound features. The new strategy is to directly detect sub-patterns from short
term power spectrum, which is relatively lower-level and at the same time more
efficient representation of the raw signals, instead of usually a few dozens (or
maximally hundreds) of features, most often used for instrument recognition
purposes. Next, we choose the multiple candidates from each frame during
the frame-wise classification based on similarity of the spectrum, and weight
them based on their possibility over all the sound period to get the more
accurate estimation of multiple instruments which are playing simultaneously
in the music piece. This approach also avoids extracting more compact feature
patterns of multiple instruments from polyphonic sounds, which is difficult and
inaccurate because of the information-loss during the abstraction process. Our
experiments show that the sub-patterns detected from the power spectrum
slices contain sufficient information for the multiple-timbre estimation tasks
and improve the robustness of instrument identification as well.

Acknowledgments

This work was supported by the National Science Foundation under grant IIS-
0414815, and also by the Research Center of PJIIT, supported by the Polish
National Committee for Scientific Research (KBN).
We are grateful to Dr. Xin Zhang for many helpful discussions we had
with her and for the comments she made which improved the quality and
readability of the paper.

References
1. Agostini G, Longari M, Pollastri E. (2001) Content-Based Classification of Mu-
sical Instrument Timbres. International Workshop on Content-Based Multime-
dia Indexing
2. American National Standards Institute (1973) American national standard:
Psychoacoustical terminology. ANSI S3.20-1973
3. Aniola P, Lukasik E (2007) JAVA Library for Automatic Musical Instruments
Recognition. AES 122 Convention, Vienna, Austria
4. Brown JC (1999) Computer identification of musical instruments using pattern
recognition with cepstral coefficients as features. J.Acoust.Soc.Am. 105, 1933–
1941
5. Fitzgerald R, Lindsay A (2004) Tying semantic labels to computational de-
scriptors of similar timbres. Sound and Music Computing’04
6. Fujinaga I, McMillan K (2000) Real Time Recognition of Orchestral Instru-
ments. International Computer Music Conference
7. Herrera P, Amatriain X, Batlle E, Serra X (2000) Towards instrument segmen-
tation for music content description: a critical review of instrument classification
techniques. International Symposium on Music Information Retrieval ISMIR
8. ISO/IEC JTC1/SC29/WG11 (2004) MPEG-7 Overview. Available at http:
//www.chiariglione.org/mpeg/standards/mpeg-7/mpeg-7.htm
 Title Suppressed Due to Excessive Length 15

9. Kaminskyj I (2002) Multi-feature Musical Instrument Sound Classifier w/user

determined generalisation performance. Proceedings of the Australasian Com-
puter Music Association Conference ACMC, 53–62
10. Kawahara T, Lee A, Kobayashi T, Takeda K, Minematsu N, Sagayama S,
Itou K, Ito A, Yamamoto M, Yamada A, Utsuro T, K. Shikano K (2000) Free
software toolkit for Japanese large vocabulary continuous speech recognition.
Proc. Int’l Conf. on Spoken Language Processing (ICSLP) , Vol. 4, 476–479
11. Kitahara T, Goto M, Okuno H G (2005) Pitch-Dependent Identification of
Musical Instrument Sounds. Applied Intelligence 23, 267-275, Springer
12. Logan B (2000) Mel Frequency Cepstral Coefficients for Music Modeling. Pro-
ceedings of the First International Symposium on Music Information Retrieval
ISMIR 2000
13. Martin K D, Kim Y E (1998) Musical instrument identification: A pattern-
recognition approach. 136-th meeting of the Acoustical Society of America,
Norfolk, VA
14. Pollard, H.F., Jansson, E.V. (1982) A Tristimulus Method for the Specification
of Musical Timbre. Acustica, 51: 162–171
15. Ras Z, Wieczorkowska A, Lewis R, Marasek K, Zhang C, Cohen A, Kolczynska
E, Jiang M (2008) Automatic Indexing of Audio With Timbre Information for
Musical Instruments of Definite Pitch, http://www.mir.uncc.edu/
16. Ras Z, Zhang X, Lewis R (2007) MIRAI: Multi-hierarchical, FS-tree based Mu-
sic Information Retrieval System (Invited Paper). In: Kryszkiewicz M, Peters
J, Rybinski H, Skowron A (eds), Rough Sets and Intelligent System Paradigms,
International Conference RSEISP 2007, Proceedings. Springer, LNAI 4585, 80–
89
17. Saha G, Yadhunandan U (2004) Modified Mel-Frequency Cepstral Coefficient.
Proceedings of the IASTED 2004
18. Sonic Foundry (2003) Sound Forge. Software
19. Wieczorkowska A (2000) Towards Musical Data Classification via Wavelet
Analysis. In: Ras, Z.W., Ohsuga, S. (eds.): Foundations of Intelligent Systems.
Proc. ISMIS’00, Charlotte, NC, USA, LNCS/LNAI, Vol. 1932, Springer-Verlag
292–300
20. Wieczorkowska A, Ras Z, Zhang X, Lewis R (2007) Multi-way Hierarchic Clas-
sification of Musical Instrument Sounds. In: S. Kim S, Park J, Pissinou N, Kim
T, Fang W, Slezak D, Arabnia H, Howard D (eds) International Conference
on Multimedia and Ubiquitous Engineering MUE 2007, Seoul, Korea. IEEE
Computer Society, Los Alamitos Washington Tokyo
21. Wold E, Blum T, Keislar D, Wheaten, J (1996) Content-based classification,
search, and retrieval of audio. Multimedia, IEEE 3(3): 27–36
22. Zhang X (2007) Cooperative Music Retrieval Based on Automatic Indexing of
Music by Instruments and Their Types. PhD dissertation, The University of
North Carolina at Charlotte, Charlotte