Merged Pitch Histograms and Pitch-duration Histograms

Hui Liu

, Tingting Xue

and Tanja Schultz

Cognitive Systems Lab, University of Bremen, Germany

Keywords:

Pitch Statistics, Pitch Histogram, Merged Pitch Histogram, Pitch-duration Histogram, Pitch-related Features,

Music Computing, Music Information Retrieval.

Abstract:

The traditional pitch histogram and various features extracted from it play a pivotal role in music information

retrieval. In the research on songs, especially applying pitch statistics to investigate the main melody, we found

that the pitch histogram may not necessarily reﬂect the notes’ pitch characteristic of the whole song perfectly.

Therefore, we took the note duration into account to propose two advanced versions of pitch histograms and

validated their applicability. This paper introduces these two novel histograms: the merged pitch histogram

by merging consecutively repeated pitches and the pitch-duration histogram by utilizing each pitch’s duration

information. Complemented by the description of their calculation algorithms, the discussion of their advan-

tages and limitations, the analysis of their application to songs from various languages and cultures, and the

demonstration of their use cases in state-of-the-art research works, the proposed histograms’ characteristics

and usefulness are intuitively revealed.

1 INTRODUCTION

In (Tzanetakis et al., 2003), the authors ﬁrst propose

the concept of pitch histogram that reﬂects the pitch

content of notes in music pieces. Due to the normal-

ization, pitch histogram is also called pitch-frequency

histogram (Gedik and Bozkurt, 2010), but the essence

is the same. A variation of the basic pitch histogram

based on music theory is the pitch class histogram,

which counts only twelve pitch classes, and the var-

ious octaves of each pitch are grouped into the same

bin. Derived from the pitch class histogram, a folded

ﬁfths pitch class histogram is calculated by reorder-

ing the bins of the original unordered histogram such

that perfect ﬁfths rather than semitones separate ad-

jacent bins. The pitch histogram also has successors

for different occasions, such as a melodic interval his-

togram that counts the distances between two consec-

utive pitches rather than the individual pitches them-

selves. A variety of features can be extracted from

the pitch histogram and its derivatives, such as range,

mean, variability, skewness, and kurtosis (McKay,

2010), applied to various machine learning-based mu-

sic research aspects like genre classiﬁcation and cul-

tural diversity analysis.

https://orcid.org/0000-0002-6850-9570

https://orcid.org/0000-0002-5815-7217

https://orcid.org/0000-0002-9809-7028

As another musical characteristic alongside the

pitch, the note duration and the rhythm resulting from

it are also important research material. A rhythmic

value histogram is a normalized histogram, where the

value of each bin speciﬁes the fraction of all notes

in the piece with a quantized rhythmic value corre-

sponding to that of the given bin (McKay, 2010).

Not similarly, a beat histogram, ﬁrst applied to Mu-

sic Information Retrieval (MIR) research in (Brown,

1993), emphasizes note onsets rather than durations,

and is also utilized for genre classiﬁcation (Tzanetakis

et al., 2001) (Tzanetakis and Cook, 2002) (Tzane-

takis, 2002) (Lykartsis and Lerch, 2015). From a

simplistic perspective, rhythm can be perceived as the

number of notes played at a speciﬁc tempo within a

bin, initiating a (normalized) duration histogram in

(Karydis, 2006).

Although the pitch sequence and the duration se-

quence of music are widely used to generate various

kinds of histograms and features, each note’s pitch

and duration information are rarely used together. For

example, (Karydis, 2006) used both pitch histograms

and (normalized) note duration histograms to gener-

ate their series of features, respectively, for the ac-

curacy improvement of symbolic music genre clas-

siﬁcation, but each note’s pitch and duration infor-

mation was not co-analyzed. As stated in (Adams

et al., 2004), while a transcription into a sequence

Liu, H., Xue, T. and Schultz, T.

Merged Pitch Histograms and Pitch-duration Histograms.

DOI: 10.5220/0011310300003289

In Proceedings of the 19th International Conference on Signal Processing and Multimedia Applications (SIGMAP 2022), pages 32-39

ISBN: 978-989-758-591-3; ISSN: 2184-9471

of (pitch, duration) pairs is convenient and musically

intuitive, no evidence shows that it is an optimal rep-

resentation (at that time).

In the study of songs’ main melody scores, we

found that the pitch histogram and the features it pro-

duces do n¡t necessarily depict the pitch distribution

of the whole song perfectly. For instance, the center

pitch of the whole song’s main melody, also called the

melody center in the automatic singing key estimation

research (Liu et al., 2022b), is often not perfectly re-

ﬂected by the mean pitch or the median pitch features

extracted from the pitch histogram. More reasonable

descriptions can be achieved by applying each note’s

pitch and duration information together. Another typ-

ical example is that jointly considering consecutive

identical pitches can, to some extent, provide a bet-

ter representation of the particular pitch distribution

of traditional Han anhemitonic pentatonic folk songs

(Liu et al., 2022a). Based on a series of experimental

veriﬁcation, we put forward and instantiate two novel

types of pitch histograms, the merged pitch histogram

and the pitch-duration histogram, combined with note

duration in some or full measure, hoping that they can

beneﬁt various aspects of music research.

The vertical axes of the three pitch histograms (ba-

sic, merged, pitch-duration) in all ﬁgures of this paper

are normalized, respectively, for a better comparison.

As introduced above, because of the normalization,

pitch histograms can be called pitch-frequency his-

tograms (Gedik and Bozkurt, 2010). Therefore, with

normalization, the two proposed histograms can be

similarly named merged pitch-frequency histograms

and pitch-duration-frequency histograms.

The instances illustrated in this paper apply the

notation-based symbolic pitch statistics instead of the

calculation of audio data. For the latter, (Tolonen and

Karjalainen, 2000) proposed a multiple pitch detec-

tion algorithm, on which the two novel histograms

proposed in this paper are also applicable without ob-

stacles.

2 TWO NOVEL VARIANTS OF

PITCH HISTOGRAMS

2.1 Merged Pitch Histogram

Consecutive notes of the same pitch in instrumen-

tal performance should be counted reasonably repeat-

edly, as the player needs to use the organ repeatedly

and proﬁciently to play these notes. For the perfor-

mance or research of songs, whether the pitch’s con-

tinuous repetition should be counted repeatedly de-

serves further investigation.

Figure 1: The ﬁrst phrase of The Sound of Silence’s main

melody.

For songs in polysyllabic languages, a word is

sometimes split into consecutive identical pitches.

The words “hello” and ”darkness” in the ﬁrst phase

of The Sound of Silence are examples (see Figure 1).

However, this is not the case for songs in monosyl-

labic languages. Therefore, it may cause problems

in calculating the pitch histogram-based features for

a music research work containing songs of both lan-

guage types.

In addition, the basic pitch histogram is not

well compatible with differences in melodic details,

including symbolic-symbolic, symbolic-audio, and

audio-audio divergences, which will be detailed in

Section 4.1.

As a modiﬁcation of the basic pitch histogram that

considers pitches’ temporal information in a simple

way, a merged pitch histogram counts consecutive

notes of the same pitch only once. If there is a rest

among a series of consecutive identical pitches, the

ﬁrst pitch after the rest should be counted again.

The red histogram in Figure 2 illustrates an exam-

ple of a merged pitch histogram representing the ﬁrst

phrase of the song That’s Why (You Go Away).

2.2 Pitch-Duration Histogram

The merged pitch histogram considers pitches’ tem-

poral information in some measure (from the perspec-

tive of continuity) to make decrements on the cor-

responding bins; In contrast, the pitch-duration his-

togram makes information increments on the bins uti-

lizing each pitch’s temporal information in full mea-

sures (from the perspective of weighting). Each bin in

the pitch-duration histogram responds to the total du-

ration of the corresponding pitch in the music piece,

instead of its number of occurrences.

The green histogram in Figure 2 displays a pitch-

duration histogram containing only the ﬁrst phrase in

the song That’s Why (You Go Away).

2.3 Comparison

Figure 2 clearly demonstrates the differences between

the two proposed novel histograms and the basic pitch

histograms:

Merged Pitch Histograms and Pitch-duration Histograms

Ba by won’t you tell me why,

Figure 2: The basic pitch histogram and its two novel vari-

ants for the ﬁrst phrase of That’s Why (You Go Away)’s main

melody. Colored down arrows mark the mean pitch features

(“melody centers”) for each type of histograms.

• The basic pitch histogram in blue emphasizes the

note C5 due to considering only the number of oc-

currences. The corresponding mean pitch (A4#) is

relatively high, close to C5.

• By merging consecutive identical pitches, the

merged pitch histogram in red presents the phrase

as an even distribution, while its mean pitch (G4)

is skewed towards the lower two pitches.

• The pitch-duration histogram in green elevates the

importance of F4# to the same level as C5 regard-

ing the entire note duration. The calculated mean

pitch (A4) is between the above two.

As a side note, the median pitch calculated by

the average of the entire phrase’s highest and lowest

pitches is G4#, different from all the mean pitches.

2.4 Histogram Calculation

For the calculation of the basic symbolic pitch his-

togram using notation-based formats, as (Tzanetakis

et al., 2003) explains, the algorithm increments the

corresponding pitch’s counter. The value in each his-

togram bin is normalized in the last stage of the cal-

culation by dividing it by the total number of pitches

of the whole piece, in order to account the variabil-

ity in the average number of pitches per unit time be-

tween different pieces of music. Because of the nor-

malization step, in some literature, like (Gedik and

Bozkurt, 2010), a pitch histogram is also called a

pitch-frequency histogram.

The merged pitch histogram is generated based

on the above basic pitch histogram calculation with

a slight modiﬁcation: if a series of consecutive notes

have an identical pitch, and there are no rests between

them, the corresponding pitch’s counter is added only

Figure 3: The basic pitch histogram and the merged pitch

histogram for the ﬁrst phrase of That’s Why (You Go

Away)’s main melody without normalization.

by one. Such a calculation is straightforward to

implement with a standard loop: the current note’s

pitch is compared with the previous one (a rest is

also counted as a note, but only for comparison). If

the same, the pitch corresponding to the note is not

counted; otherwise, add one to the pitch’s counter.

Note that the denominator in the ﬁnal normalization

process is not the same as in the basic pitch histogram.

It is worth pointing out that without normalization,

we can directly observe which pitches are merged

and how much is the quantity of notes being merged

by comparing the merged pitch histogram to the ba-

sic pitch histogram, as Figure 3 evidences. After

normalization, the conclusion drawn by height com-

parison changes slightly. First, we exclude the easy

case where no merging happens in the whole melody.

The pitch bins that do not have merged notes must

be higher in the normalized merged pitch histogram

than in the normalized basic pitch histogram (see the

pitches E4 and F4# in Figure 2) due to the fact that the

numerators of the two calculations are identical and

the former’s denominator is less than the latter’s. The

contrapositive conclusion is that a pitch must con-

tain merged notes if its bin’s normalized height in

the merged pitch histogram is shorter than or equal to

its corresponding normalized height in the basic pitch

histogram (see the pitch C5 in Figure 2).

The construction process of the pitch-duration his-

togram can go simply through a traversal of the notes

in the piece. For each note, its duration is accumu-

lated over the bin of the corresponding pitch. After

processing all notes, the histogram is normalized by

dividing by the sum of the durations of all pitches.

3 INSTANCES AND ANALYSIS

For intuitive analysis, we apply the basic pitch his-

togram, the merged pitch histogram, and the pitch-

duration histogram to a set of folk songs from differ-

SIGMAP 2022 - 19th International Conference on Signal Processing and Multimedia Applications

Table 1: Information on all songs analyzed in this paper.

Song Lyricist Composer Language/Country Pitch range

The Sound of Silence Paul Simon English/USA 18 semitones

That’s Why (You Go Away) Jascha Richter English/Denmark 21 semitones

Auld Lang Syne Robert Burns Scottish/UK 18 semitones

Vo Luzern/Luz

arn auf/uf W

aggis zue Johann L

uthi German/Switzerland 18 semitones

Bella Ciao Italian 14 semitones

Arirang Korean 13 semitones

Anile Anile Vaa Vaa Vaa Tamil/India 13 semitones

Troika Pochtovaya Russian 16 semitones

Sakura Sakura Japanese 14 semitones

The Green Poplar and Willow Mandarin 15 semitones

Figure 4: Three types of pitch histograms for Bella ciao’s

main melody with Movable-Do. The score segment below

indicates the phrases where E4 and F4 are repeated.

ent languages and cultures. Since most of them do

not have original deﬁnite keys, they are all notated

using Movable-Do. Table 1 lists the information on

the songs, including the three exempliﬁed in Sections

2 and 4.

Bella Ciao

The merged pitch histogram of the Italian folk song

Bella Ciao illustrated in Figure 4 has shorter bins for

E4 and F4 (in Movable-Do) compared to the basic

pitch histogram, resulting in correspondingly longer

bins for all other pitches. Such a phenomenon mir-

rors the fact that only E4 and F4 are repeated succes-

sively in Bella Ciao’s main melody, the former four

times and the latter twice (see the score segment be-

low in Figure 4). In contrast, the importance of E4 and

F4, together with A3, is lifted in the pitch-duration

histogram, highlighting the fact that these pitches are

sung for extended periods throughout the whole song.

It is well understood that the pitch-duration his-

togram generates the same mean pitch (C4) as the ba-

sic pitch histogram, while the merged pitch histogram

gives one-semitone lower mean pitch (B3). The me-

dian pitch, generated by the average of the highest and

Figure 5: Three types of pitch histograms for Arirang’s

main melody with Movable-Do.

lowest pitches of the song, lies between A3# and B3

and is evidently not comprehensive in reﬂecting the

melody center.

Arirang

As shown in Figure 5, the mergeability of the pitches

C4 and G4 (Movable-Do) is witnessed in the Korean

anhemitonic pentatonic folk song Arirang. Mean-

while, the pitch-duration histogram exchanges the im-

portance of G3 and A3 while emphasizing C4’s sig-

niﬁcance, strongly reﬂecting this song’s major penta-

tonic scale mode.

Auld Lang Syne

Among the main melody histograms for the Scot-

tish ballad Auld Lang Syne exhibited in Figure 6, the

merged pitch histogram reverses the importance of

G3 and A3 compared to the basic pitch histogram,

while the pitch-duration histogram acts signiﬁcantly

on A4. This anhemitonic pentatonic song’s basic and

merged histograms have been employed recently in

a new research topic of computational ethnomusicol-

ogy, of which more introduction will be provided in

Section 4.2.

Merged Pitch Histograms and Pitch-duration Histograms

Figure 6: Three types of pitch histograms for Auld Lang

Syne’s main melody with Movable-Do.

More Instances

Further examples of folk songs in various languages

and cultures are given in Figure 7. Most folk songs

lack the original scores and are more likely to have

divergent details of symbolic melodies than modern

songs, which will be detailed in Section 4.1. Apply-

ing the merged pitch histogram and the pitch-duration

histogram will help dilute these divergences’ effects.

Figure 7: Three types of pitch histograms for Anile Anile

Vaa Vaa Vaa (top), Troika Pochtovaya (middle), and Sakura

Sakura (bottom) with Movable-Do.

4 DISCUSSION AND USE CASES

4.1 Advantages and Limitations

The potential advantages of merged pitch histograms

and pitch-duration histograms are (1) bridging the in-

ﬂuence of different linguistic characteristics, which is

already explained in Section 2.1; (2) helping down-

play differences in melodic details of a song’s various

versions, which, in MIR studies, can be divided into

three types of divergence:

• Symbolic-symbolic divergences usually occur

when the song or music does not have an original

score. Many traditional folk songs have been cre-

ated by local people’s improvisation and handed

down orally among the people, then notated by

several individuals. Hence, the scores of each ver-

sion may differ in many details.

• Singers sing, and musicians play the same piece

with their characteristics, leading to symbolic-

audio divergences.

• For the same reasons as the previous one, the

melodies that different singers and musicians

eventually produce differ from each other, caus-

ing audio-audio divergences.

A typical instance is the yodels, where it can be

noticed that many songs’ scores differ in melodic de-

tails (e.g., whether some parts use syncopation or not)

in different collections, causing a symbolic-symbolic

divergence. When sung, the yodeling part often uses

a series of successive identical pitches, which usually

do not appear on the score, creating a symbolic-audio

divergence. The audio-audio divergence is easy to

imagine and does not need to be elaborated on. Both

merged pitch histograms and pitch-duration pitch his-

tograms eliminate the effect that makes the yodeling

pitches’ bins in unnecessarily high counts produced

by repeatedly being sung, thus providing a reasonable

de-peaking of the song’s overall pitch distribution.

Figure 8 conveys an example of the symbolic-

symbolic divergence (marked in red). Its symbolic-

audio divergence and audio-audio divergence can be

reﬂected in each singer’s yodeling parts. All the di-

vergences can be addressed by the merged pitch his-

togram and the pitch-duration histogram.

The pitch-duration histogram is an improvement

that reﬂects as comprehensively as possible the im-

portance of each pitch in the piece. However, from a

melodic divergence perspective, the basic pitch his-

togram and the merged pitch histogram are some-

times more advantageous for pitch statistics than the

pitch-duration histogram, especially in cases where

the pitch duration can be arbitrarily prolonged. For

SIGMAP 2022 - 19th International Conference on Signal Processing and Multimedia Applications

Figure 8: Top: one version of the scores for the Swiss yo-

del Vo Luzern/Luz

arn auf/uf W

aggis zue generated using

the XML ﬁle downloaded from the Alojado Lieder Archiv

website

. The score fragment below shows the divergence

marked in red that often occurs in other sources. The

merged pitch histogram can eliminate the divergence.

example, for the same position of one phrase’s end,

some scores extend the beat to the end of the bar,

while some give a rest. The basic and the merged

pitch histograms count such a pitch once, making

more sense than fully considering its duration in the

pitch-duration histogram.

4.2 Progressing Cultural Diversity in

Computational Ethnomusicology/

MIR with Merged Pitch Histograms

Recent studies have found that the vast majority of

traditional Han anhemitonic pentatonic folk songs can

be identiﬁed intuitively according to their distinctive

bell-shaped pitch distribution in pitch histograms, re-

ﬂecting the Chinese characteristics of Zhongyong (the

doctrine of the mean) and following the trend from

an ethnocultural perspective (Liu et al., 2022a). The

bell-shaped basic/merged pitch histograms are also

witnessed in some anhemitonic pentatonic folk songs

of other East Asian ethnic groups, such as Japanese

and Korean (exempliﬁed in Figure 5), indicating the

exchange between countries, peoples, and cultures

throughout history, although they also have unique

musical systems of their own.

As one of the study’s results, the bins produced in

the basic pitch histogram that violate the bell shape

can be improved by the merged pitch histogram to a

great extent.

Take Figure 9 as an example. This song’s basic

pitch histogram generates signiﬁcant exceptions at the

lowest pitch that violate the bell-shaped curve. It is

interesting to note that the “Do Do Do” is frequently

repeated at several phrase ends of this song to ex-

https://www.lieder-archiv.de/

Figure 9: Three types of pitch histograms for The Green

Poplar and Willow’s main melody with Movable-Do.

press an upbeat rhythm, causing this pitch to have a

higher count number (Liu et al., 2022a). By utilizing

the merged pitch histogram, the pitch statistics show

a perfect bell shape.

For comparison, the Scottish folk song Auld Lang

Syne is also anhemitonic pentatonic (see Figure 6),

but its pitch histogram is far from perfect bell-shaped.

Its merged pitch histogram is even less bell-shaped.

The merged pitch histogram were applied together

with the anhemitonic pentatonic pitch histogram pro-

posed in (Liu et al., 2022a) to create a merged anhemi-

tonic pentatonic pitch histogram, playing a helpful

role in the feature design and genre classiﬁcation re-

search. In addition, an existing feature from the Time

Series Feature Extraction Library (TSFEL) (Baran-

das et al., 2020), Negative Turning, and a designed

novel feature, Degree of Bell Shape (DoBell), are ex-

tracted from various types of histograms to describe

to which extent a pitch distribution approximates a

bell shape. The preliminary classiﬁers built with these

features performed well, indicating that lightweight

machine learning applying only pitch histograms and

merged pitch histograms can promote cultural diver-

sity in MIR.

4.3 Automatic Singing Key Estimation

Applying New Histogram Variants

The mean pitch features derived from the three types

of histograms involved in this paper play an essential

role in the recent study of automatic singing key esti-

mation (Liu et al., 2022b).

Compared to the simple median pitch feature that

is calculated as the average of the lowest and highest

pitches, the mean pitch features generated from the

three pitch histograms are more amenable for suggest-

ing the appropriate keys for individuals, among which

the mean pitch of the pitch-duration histogram is the

most reliable.

Figure 10 gives a running interface screenshot

of the software Automatic Singing Key Estimator

(ASKE, pronounced as “Ask-key”) presented in (Liu

Merged Pitch Histograms and Pitch-duration Histograms

Figure 10: Screenshot of the software Automatic Singing

Key Estimator (ASKE).

et al., 2022b). For Auld Lang Syne, only one key was

automatically calculated for each amateur singer us-

ing a merged pitch histogram or a pitch-duration his-

togram, and the resulted key was scored satisfactorily

by the subjects in the singing experiment.

5 CONCLUSION AND OUTLOOK

This paper introduces two novel variants of the basic

pitch histogram, the merged pitch histogram and the

pitch-duration histogram, which consider each pitch’s

temporal information in the entire melody, the for-

mer in some measure (from the perspective of con-

tinuity) and the latter in full measure (from the per-

spective of weighting). Complemented by the anal-

ysis of song examples and the comparison between

histograms, the characteristics of the proposed his-

tograms are visually expounded. Furthermore, the

proposed histograms’ computational algorithms, ad-

vantages, limitations, and use cases in the latest re-

search works are exhibited in detail.

We believe that the merged pitch histogram and

the pitch-duration histogram are meaningful and help-

ful for music research, especially for the informa-

tion retrieval of songs’ main melody, as exempliﬁed

in Sections 4.2 and 4.3. We expect the two intro-

duced histograms to be extensively applied in vari-

ous research aspects. Future work includes investi-

gating abundant new features extracted from the two

novel pitch histograms and applying these features to

various machine learning-based music research ﬁelds.

Besides, diverse transformation approaches of the ba-

sic pitch histogram, such as pitch class, folded ﬁfths,

and melodic interval, can be practiced on the two

novel variants to generate more varied histograms

and new features, like the kurtosis of the pitch-class-

duration histogram.

Moving from overall pitch statistics to segment-

based features, windowing/framing the song’s

(pair, duration) sequence to generate time series

of merged pitch histograms and pitch-duration

histograms deserves further investigating, e.g., uti-

lizing the Time Series Subsequence Search Library

(TSSEARCH) (Folgado et al., 2022) to conduct a

subsequence similarity analysis on the melody.

REFERENCES

Adams, N. H., Bartsch, M. A., Shifrin, J., and Wakeﬁeld,

G. H. (2004). Time series alignment for music infor-

mation retrieval. In ISMIR 2004 – 5th International

Conference on Music Information Retrieval.

Barandas, M., Folgado, D., Fernandes, L., Santos, S.,

Abreu, M., Bota, P., Liu, H., Schultz, T., and Gam-

boa, H. (2020). Tsfel: Time series feature extraction

library. SoftwareX, 11:100456.

Brown, J. C. (1993). Determination of the meter of musical

scores by autocorrelation. The Journal of the Acousti-

cal Society of America, 94(4):1953–1957.

Folgado, D., Barandas, M., Antunes, M., Nunes, M. L.,

Liu, H., Hartmann, Y., Schultz, T., and Gamboa, H.

(2022). Tssearch: Time series subsequence search li-

brary. SoftwareX, 18:101049.

Gedik, A. C. and Bozkurt, B. (2010). Pitch-frequency

histogram-based music information retrieval for turk-

ish music. Signal Processing, 90(4):1049–1063.

Karydis, I. (2006). Symbolic music genre classiﬁcation

based on note pitch and duration. In ADBIS 2006

– 10th East European Conference on Advances in

Databases and Information Systems, pages 329–338.

Springer.

Liu, H., Jiang, K., Gamboa, H., Xue, T., and Schultz, T.

(2022a). Bell shapes: Exploring the pitch distribution

of traditional han anhemitonic pentatonic folk songs.

(submitted).

Liu, H., Xue, T., Gamboa, H., Hertenstein, V., Xu, P., and

Schultz, T. (2022b). Automatic singing key estimation

for amateur singers using novel pitch-related features:

A singer-song-feature model. (submitted).

Lykartsis, A. and Lerch, A. (2015). Beat histogram features

for rhythm-based musical genre classiﬁcation using

multiple novelty functions. In DAFx15 – 18th Inter-

national Conference on Digital Audio Effects.

McKay, C. (2010). Automatic music classiﬁcation with

jMIR. PhD thesis, McGill University.

Tolonen, T. and Karjalainen, M. (2000). A computationally

efﬁcient multipitch analysis model. IEEE transactions

on speech and audio processing, 8(6):708–716.

Tzanetakis, G. (2002). Manipulation, analysis and retrieval

systems for audio signals. Princeton University.

Tzanetakis, G. and Cook, P. (2002). Musical genre classiﬁ-

cation of audio signals. IEEE Transactions on speech

and audio processing, 10(5):293–302.

SIGMAP 2022 - 19th International Conference on Signal Processing and Multimedia Applications

Tzanetakis, G., Ermolinskyi, A., and Cook, P. (2003). Pitch

histograms in audio and symbolic music information

retrieval. Journal of New Music Research, 32(2):143–

152.

Tzanetakis, G., Essl, G., and Cook, P. (2001). Automatic

musical genre classiﬁcation of audio signals. In Pro-

ceedings of the 2nd international symposium on music

information retrieval, Indiana, volume 144.

Merged Pitch Histograms and Pitch-duration Histograms