- require "base64"
~ "\xEF\xBB\xBF"
- def quellen opts
- etc = opts.key? :etc
- if etc
- opts.delete :etc
- etc = "\n"
- else
- etc = ''
- "#{opts.map {|k, v| "#{k}" }.join "\n"}#{etc}"
- def link link
- "#{link}"
- def import_data file
- mime_type = IO.popen(["file", "--brief", "--mime-type", file], in: :close, err: :close) { |io| io.read.chomp }
- content = Base64.urlsafe_encode64 File.read( file)
- "data:#{mime_type};base64,#{content}"
!!! 5
%html(lang='en')
%head
-#%meta(charset="utf-8")
%title Decoding the sound of 'hardness' and 'darkness' as perceptual dimensions of music
-#%link(rel="stylesheet" href="fonts/Roboto.css")
-#%link(rel="stylesheet" href="fonts/RobotoSlab.css")
-#%link(rel="stylesheet" href="fonts/PT_Mono.css")
-#%link(rel="stylesheet" href="fonts/PT_Sans.css")
-#%link(rel="stylesheet" href="fonts/Vollkorn.css")
-#%link(rel="stylesheet" href="fonts/Asset.css")
-#%link(rel="stylesheet" href="fonts/WithinDestruction.css")
-#%link(rel="stylesheet" href="fonts/BlackDahlia.css")
-#%link(rel="stylesheet" href="fonts/ThroughStruggleDEMO.css")
-#%link(rel="stylesheet" href="fonts/TheDefiler.css")
%link(rel="stylesheet" href="fonts/Cardo.css")
%link(rel="stylesheet" href="fonts/Italianno.css")
-#%link(rel="stylesheet" href="fonts/CinzelDecorative.css")
%link(rel="stylesheet" href="style.css")
%meta(name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=no")
%body
%header(style="")
%figure.logos(style="margin-top:0.3cm")<>
%img#tagungs-logo(style="float:right" src="files/icmpc15_logo.jpg")
%img#uni-logo(src="files/univie_logo.png")
-#%div(style="font-size:0.8em;margin-top:1.31cm")
44. Jahrestagung für Akustik
%br<>
Technische Universität München
%br<>
19. März 2018 .. 22. März 2018
-#.grabstein
.grabstein-was DAGA
.grabstein-wo Technische Universität München
.grabstein-von ✦ 19. März 2018
.grabstein-bis ✝ 22. März 2018
-#%img(style="height:7cm;top:3cm;right:24cm;position:absolute" alt="Dunkle Nacht" src="files/Candle.png")
%h1
Decoding the sound of hardness
and darkness
as perceptual dimensions of music
%p#authors<>
%span.author(data-mark="1,2")<> Isabella Czedik-Eysenberg
%span.author(data-mark="1")<> Christoph Reuter
%span.author(data-mark="2")<> Denis Knauf
%p#institutions<>
%span.institution(data-mark="1")<> University of Vienna, Austria
%span.institution(data-mark="2")<> Student at Technical University of Vienna, Austria
%main
#column1_1
%section#heavy_features
:markdown
Sound Features
==============
Considering Bonferroni correction, 65 significant feature
correlations were found for the concept of hardness
.
The characterizing attributes of hardness
include high
tempo and sound density, less focus on clear melodic lines than
noise-like sounds and especially the occurrence of strong percussive
components.
%ol
%li
percussive energy / rhythmic density
%figure
%img(style="width:50%" src="files/sonagramm_blunt_log.png")
%img(style="width:50%" src="files/sonagramm_decap_log.png")
%li
dynamic distribution
%figure
%img(style="width:50%" src="files/blunt_envelope.png")
%img(style="width:50%" src="files/decap_envelope.png")
%figure
%img(style="width:50%" src="files/blunt_dyndist.png")
%img(style="width:50%" src="files/decap_dyndist.png")
%li
melodic content / harmonic entropy
%figure
%img(style="width:50%" src="files/blunt_chromagram.png")
%img(style="width:50%" src="files/decap_chromagram.png")
%section#heavy_model
%h1 Model
:markdown
Sequential feature selection
* set of 5 features
* predictive linear regression model
RMSE | 0.64
R-Squared | 0.80
MSE | 0.40
MAE | 0.49
r | 0.900
%figure
%img(src="scatter_hardness_model5.png")
%section#heavy_rater_agreement
:markdown
Rater Agreement
===============
Intraclass Correlation Coefficient (Two-Way Model, Consistency): 0.653
#column1_2
-#%section#aims
%h1 Aims
%p
Based on computationally obtainable signal features, the creation
of models for the perceptual concepts of hardness
and
darkness
in music is aimed for. Furthermore it shall be
explored if there are interactions between the two factors and to
which extent it is possible to classify musical genres based on
these dimensions.
%section#method
%h1 Method
%figure.right(style="width:50%")
%img(src="files/LastFM.png")
:markdown
Based on last.fm listener statistics, 150 pieces of music were selected
from 10 different subgenres of metal, techno, gothic and pop music.
In an online listening test, 40 participants were asked to rate the
refrain of each example in terms of hardness
and darkness
.
These ratings served as a ground truth for examining the two
concepts using a machine learning approach:
Taking into account 230 features describing spectral distribution,
temporal and dynamic properties, relevant dimensions were
investigated and combined into models.
Predictors were trained using five-fold cross-validation.
%figure.right(style="width:50%")
%img(src="files/einhorn/diagramm_vorgang_english.png")
%section#data
%h1 Data
%figure.right(style="width:50%")
%img(src="files/scatter_hard_dark_dashedline_2017-09-05.png")
%section#hardness
%h1 Hardness
%p
Hardness
is often considered a distinctive feature of (heavy)
metal music, as well as in genres like hardcore techno or Neue
Deutsche Härte
.
In a previous investigation the concept of hardness
in music
was examined in terms of its acoustic correlates and suitability as
a descriptor for music #{quellen 'Czedik-Eysenberg et al.' => 2017}.
#column1_3
%section#darkness
%h1 Darkness
%p
Certain kinds of music are sometimes described as dark
in a
metaphorical sense, especially in genres like gothic or doom metal.
According to musical adjective classifications dark
is part
of the same cluster as gloomy
, sad
or
depressing
#{quellen Hevner: 1936}, which was later adopted in
computational musical affect detection
#{quellen 'Li & Oghihara' => 2003}.
This would suggest the
relevance of sound attributes that correspond with the expression
of sadness, e.g. lower pitch, small pitch movement and dark
timbre #{quellen Huron: 2008}. In timbre research brightness
is often considered one of the central perceptual axes
#{quellen Grey: 1975, 'Siddiq et al.' => 2014}, which raises the
question if darkness
in music is also reflected as the
inverse of this timbral brightness
concept.
%section#darkness_features
:markdown
Sound Features
==============
Considering Bonferroni correction, 35 significant feature
correlations were found for the darkness
ratings.
While a suspected negative correlation with **timbral
brightness
** cannot be confirmed, darkness
appears to
be associated with a high **spectral complexity** and harmonic
traits like **major or minor mode**.
%figure
%img(src="files/scatter_spectral_centroid_essentia_darkness.png")
:markdown
Correlations between darkness rating and measures for brightness:
Feature | r | p
-----------------------|--------|----------
Spectral centroid | 0.3340 | <0.01
High frequency content | 0.1526 | 0.0631
%figure
%img(src="files/violin_keyEdma_darkMean_blaugelb.png")
%p
Musical excerpts in minor mode were significantly rated as
harder
than those in major mode. (p < 0.01
according to t-test)
%section#darkness_model
%h1 Model
%figure
%img(src="files/scatter_darkness_model8.png")
:markdown
Sequential feature selection:
* combination of 8 features
* predictive linear regression model
RMSE| 0.81
R-Squared| 0.60
MSE| 0.65
MAE| 0.64
r| 0.7978
%section#darkness_rater_agreement
:markdown
Rater Agreement
===============
Intraclass Correlation Coefficient (Two-Way Model, Consistency):
**0.498**
%footer
%section#further_resultes_conclusion
:markdown
Further Results & Conclusions
=================================
Comparison
----------
When comparing darkness
and hardness
, the results
indicate that the latter concept can be more efficiently described
and modeled by specific sound attributes:
* The consistency between ratings given by different raters is
higher for hardness
(see Intraclass Correlation
Coefficients)
* For the hardness
dimension, a model can be based on a more
compact set of features and at the same time leads to a better
prediction rate
Further application
-------------------
Although a considerable linear relation
(r = 0.65, p < 0.01) is present between
the two dimensions within the studied dataset, the concepts prove to
be useful criteria for distinguishing music examples from different
genres.
E.g. a simple tree can be constructed for classification into broad
genre categories (Pop, Techno, Metal, Gothic) with an accuracy of
74%.
%img(src="files/predictionTree_genreAgg2.png")
%img(src="files/confusionMatrix_simpleTree_genreAgg2.png")
%section#conclusion
:markdown
Conclusion
==========
Hardness
and darkness
constitute perceptually relevant
dimensions for a high-level description of music. By decoding the
sound characteristics associated with these concepts, they can be
used for analyzing and indexing music collections and e.g. in a
decision tree for automatic genre prediction.
-#%section#ergebnisse1(style="height:96.35cm")
%h1 4. Ergebnisse
%figure.right(style="width:70%")
%img(alt='Verwelkter Mohn' src='files/violin_genre_darkMean.svg')
%p
Es zeigt sich ein Bezug zwischen dem Genre und der
durchschnittlichen Düsterkeitsbewertung der jeweiligen Stimuli.
%figure.right(style="width:35%")
%img(alt='Ernstes Indigo' src='files/scatter_spectral_centroid_essentia_darkness.svg')
%p
Eine Antiproportionalität zu klangfarblicher Helligkeit
lässt
sich (mit der vorliegenden Messmethode) nicht nachweisen. Es liegt
im Gegenteil sogar eine leicht positive Korrelation vor –
womöglich u.a. bedingt durch erhöhte dissonante Klanganteile im
Hochfrequenzbereich (z.B. Schlagzeugvorkommen). Werden die
perkussiven Signalanteile zuvor ausgefiltert, verringert sich
dieser Effekt bereits deutlich.
%figure.nobrtd(style="width:24em")
:markdown
Merkmal|r|p
---|---|---
Spectral Centroid|0,3340|< 0,0001
Hochfrequenzanteil (> 1500 Hz)|0,1526|0,0631
Spectral Centroid (harmonischer Teil)|0,2094|0,0101
Hochfrequenzanteil (harmonischer Teil)|0,1270|0,1215
{:.merkmale}
%figcaption
Korrelation der durchschnittlichen Düsterkeitsbewertung mit Maßen
für klangfarbliche Helligkeit.
.clear
%figure.left(style="width:41.1%")
%img(alt='Trauriges Purpur' src='files/violin_keyEdma_darkMean_blaugelb.svg')
%figure
%figure.right(style="width:12em")
%img(alt="lilien grau" src="files/meanspectra_10khz_600dpi.png")
%figure.right
:markdown
Merkmal|r|p
---|---|---
RMS Gammatone 1|- 0,3989|< 0,0001
RMS Gammatone 4|- 0,3427|< 0,0001
RMS Gammatone 5|- 0,3126|0,0001
{:.merkmale}
%p(style="clear:right")
Zwischen den 30 am düstersten bzw. am wenigsten düster bewerteten
Klangbeispielen zeigen sich charakteristische Unterschiede in der spektralen
Verteilung (insbesondere im Bereich der Gammatone-Filterbank-Bänder 1, 4 und 5).
%p(style="clear:right")
Ein deutlicher Zusammenhang zeigt sich mit der Tonart der
jeweiligen Ausschnitte: Moll-Beispiele wurden im Durchschnitt als
düsterer bewertet als Stücke in Dur-Tonarten (p < 0.0001 laut t-Test).
%p(style="clear:right")
Teilweise eher statische Tonchroma-Veränderungen im Fall der als
düster bewerteten Beispiele könnten die Theorie geringere
Tonhöhenbewegungen in Zusammenhang mit einem Ausdruck von Trauer
bestätigen (siehe z.B. Chromagramm Sunn 0)))
).
%figure.right(style="width:58.2%")
%img(style="width:49%" alt='Schrumpeliges Gelb' src='files/chromagramm_sunn.svg')
%img(style="width:49%" alt='Vergängliches Weiß' src='files/chromagramm_abba.svg')
%p(style="clear:left;max-width: 50%")
Der stärkste Zusammenhang lässt sich zur Spectral Complexity
feststellen, welche die Komplexität des Signals in Bezug auf seine
Frequenzkomponenten anhand der Anzahl spektraler Peaks im Bereich
zwischen 100 Hz und 5 kHz beschreibt. Dies ist interessant mit den
Ergebnissen von #{quellen 'Laurier et al.' => 2010} in Bezug zu setzen,
welche beobachteten, dass entspannte
(relaxed
) Stücke eine
niedrigere spektrale Komplexität aufweisen, fröhliche
(happy
)
Stücke jedoch eine leicht höhere spektrale Komplexität als
nicht fröhliche
.
%figure.left(style="width:59.83%;position:relative")
%img(alt='Totes Grün' src='files/scatter_model8_mit_beschriftung_gross.svg')
%img(alt="Farbiges Beispiel" style="width:5cm;opacity:0.7;position:absolute;top:0;left:3cm" src="files/bat.png")
%p(style="clear:right")
Nach sequentieller Merkmalsauswahl wurden 8 Signaldeskriptoren zur
Bildung eines Modells zu Rate gezogen:
:markdown
Merkmal|r|p
----|----|----
Spectral Complexity (mean)| 0,6224| < 0,0001
HPCP Entropy (mean)| 0,5355| < 0,0001
Dynamic Complexity| - 0,4855| < 0,0001
Onset Rate| - 0,4837| < 0,0001
Pitch Salience| 0,4835| < 0,0001
MFCC 3 (mean)| 0,4657| < 0,0001
Spectral Centroid (mean)| 0,3340| < 0,0001
RMS Energy Gammatone 4| - 0,3427| < 0,0001
{:.merkmale}
%p
Anhand dieser wurde unter 5-facher Kreuzvalidierung ein lineares
Regressionsmodell zur Abschätzung der Düsterkeitsbewertung erstellt.
:markdown
Merkmal|Wert
----|----
Root-mean-squared error (RMSE)|0,8100
Bestimmtheitsmaß (R2)|0,6000
Mean Squared Error (MSE)|0,6500
Mean Average Error (MAE)|0,6400
Korrelation (insgesamt)|0,7978
{:.merkmale}
%div(style="clear:left")
.clear
%section#references
-#(style="width:44.5%;display:inline-block;float:right")
%h1 References
%ul.literatur
%li
%span.author Czedik-Eysenberg, I., Knauf, D., & Reuter, C.
%span.year 2017
%span.title Hardness
as a semantic audio descriptor for music using automatic feature extraction
%span.herausgeber Gesellschaft für Informatik, Bonn
%span.link
%a(href="https://doi.org/10.18420/in2017_06") https://doi.org/10.18420/in2017_06
%li
%span.author Grey, J.M.
%span.year 1975
%span.title An Exploration of Musical Timbre
%span.herausgeber Stanford University, CCRMA Report No.STAN-M-2
%li
%span.author Li,T., Ogihara,M.
%span.year 2003
%span.title Detecting emotion in music
%nobr
%span.herausgeber 4th ISMIR Washington & Baltimore
%span.pages 239-240
%li
%span.author Huron, D.
%span.year 2008
%span.title A comparison of average pitch height and interval size in major-and minor-key themes
%nobr
%span.herausgeber Empirical Musicology Review, 3
%span.pages 59-63
%li
%span.author Siddiq,S. et al.
%span.year 2014
%span.title Kein Raum für Klangfarben - Timbre Spaces im Vergleich
%nobr
%span.herausgeber 40. DAGA
%span.pages 56-57
.clear