Intelligent Audio Analysis

358 Pages·2013·5.842 MB·English

by  Björn W. Schuller (auth.)

Signals and Communication Technology
Björn W. Schuller
Intelligent
Audio Analysis
Signals and Communication Technology
For furthervolumes:
http://www.springer.com/series/4748
Björn W. Schuller
Intelligent Audio Analysis
123
BjörnW. Schuller
LSfürMensch-Maschine-Kommunikation
TU München
München
Germany
ISSN 1860-4862
ISBN 978-3-642-36805-9 ISBN 978-3-642-36806-6 (eBook)
DOI 10.1007/978-3-642-36806-6
SpringerHeidelbergNewYorkDordrechtLondon
LibraryofCongressControlNumber:2013933575
(cid:2)Springer-VerlagBerlinHeidelberg2013
Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpartof
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation,broadcasting,reproductiononmicrofilmsorinanyotherphysicalway,andtransmissionor
informationstorageandretrieval,electronicadaptation,computersoftware,orbysimilarordissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purposeofbeingenteredandexecutedonacomputersystem,forexclusiveusebythepurchaserofthe
work. Duplication of this publication or parts thereof is permitted only under the provisions of
theCopyright Law of the Publisher’s location, in its current version, and permission for use must
always be obtained from Springer. Permissions for use may be obtained through RightsLink at the
CopyrightClearanceCenter.ViolationsareliabletoprosecutionundertherespectiveCopyrightLaw.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publicationdoesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexempt
fromtherelevantprotectivelawsandregulationsandthereforefreeforgeneraluse.
While the advice and information in this book are believed to be true and accurate at the date of
publication,neithertheauthorsnortheeditorsnorthepublishercanacceptanylegalresponsibilityfor
anyerrorsoromissionsthatmaybemade.Thepublishermakesnowarranty,expressorimplied,with
respecttothematerialcontainedherein.
Printedonacid-freepaper
SpringerispartofSpringerScience+BusinessMedia(www.springer.com)
For Thorben Amadeus Bryan &
Benno Olav Sylvain
Foreword
IntelligentAudioAnalysisunitesmethodsofaudiosignalprocessingandmachine
learning. Other terms exist for this field or sub-fields and might have been used
instead,suchasComputerAuditionorMachineListening—eachofwhichisbeing
usedbypartlydifferentresearchcommunitieswithslightlydifferentunderstanding
of the core application field and the inventory of methods.
BesidesAutomaticSpeechRecognitionbeingresearchedsincemorethanhalfa
century, recently an increasing number of further speech and speaker character-
isation tasks have been pursued in the literature. In addition, the younger field of
Music Information Retrieval is growing and there is emerging interest in the
computationally ‘intelligent’ analysis of general sound events. Fields of applica-
tion comprise audio coding, edition, interaction, search, surveillance as well as
coaching and entertainment applications.
This book first propagates a unified view on the multiplicity of resulting tasks.
It further provides a broad overview of the field enriched by extensive recent
researchapplicationexamplesmostlybasedontheauthor’slatestwork.Thefocus
thereby lies on realistic conditions and standardisation by open-source software
implementations and comparative evaluations. The main goal is to increase
robustnessbytemporaryandinnovativemethodssuchasautomateddata-acquisition
by semi-supervised learning, audio signal enhancement by non-negative matrix
factorisation, systematic feature brute-forcing and application of memory-
enhanced learning algorithms—for example in combination with graphical model
structures. Machine-based recognition of speech, non-linguistic vocalisations
and para-linguistic speaker states and traits serve as examples of application in
the domain of speech processing. As for music processing, examples include
blindseparationofinstruments,determinationoftempo,metreandballroomdance
style, as well as analysis of musical key, chord progression and structure, next
toestimationofmusicmoodandsingertraits.Finally,examplesarecomplemented
bytherecognitionofgeneralsoundeventsalongwiththeiremotionalconnotation.
Inthe outlook, avenues towardsevolutionary, unsupervised andholistic audio-
signal analysis are shown.
It is thus hoped that the book may find interest by the very broad and interdis-
ciplinaryrangeofresearchersandpractitionersinacademiaandindustryreaching
fromengineeringandcomputersciencetothefieldsofspeech,language,musicand
vii
viii Foreword
generalaudiosciencewiththeirmanifoldsub-fields.Itfurtheraddresseslevelsfrom
earlytoveryadvancedlevel—obviously,though,notalldetailscanbeprovidedat
anytime,andfurtherreadingwillbeofhelpwherethereaderfindsitmosthelpful
for oneself.
Preface
This book is based on my habilitation thesis and by that on selected essential
research application examples added by explanatory chapters made during the
period of my habilitation at the Institute for Human–Machine Communication of
the Technische Universität München (TUM) in Munich, Germany, to obtain the
German state doctorate (fakultas docendi) and private lectureship (venia legendi,
German PD) in the subject area of Signal Processing and Machine Intelligence.
A representative selection of application examples was made basing on coverage
of the broader field, scientific relevance and recency. The book further includes
knowledgeandfindingsofresearchconductedandlecturesheldduringthisperiod
at TUM, the CNRS LIMSI’s Spoken Language Processing Group in Orsay,
France,theImperialCollegeLondon’sDepartmentofComputinginLondon,UK,
the Università Politecinicà delle Marche in Ancona, Italy and the National ICT
Australia in Sydney, Australia.
The aim is to provide a handbook that can be read from the beginning to the
end, structured into methods and examples of their application. Reference to the
original research is repeatedly made throughout, such that the interested reader is
referred to these, as well as to further reading from myself and my colleagues or
further research in the field. By that, the book introduces a broader view on and
newavenuestowardsthecomputationaland‘intelligent’analysisofaudioaiming
at the higher goal of lending machines the ability to listen to and understand
arbitrary and complex compounds of speech, music and sound.
Gilching, December 2012 Björn W. Schuller
ix
Acknowledgments
Great discoveries and improvements invariably involve the
co-operation of many minds. I may be given credit for
having blazed the trail, but when I look at the subsequent
developmentsIfeelthecreditisduetoothersratherthanto
myself.
—AlexanderGrahamBell
I would first like to express my deepest gratitude to my habilitation mentor and
year-long advisor Prof. Gerhard Rigoll of Technische Universität München in
Munich, Germany, for his immeasurable technical inspiration and guidance and
consultancyalongtheacademicpath.Next,fortheirutmostvaluablepreceptorship
and chance of common scientific work my further habilitation mentors Prof.
Kristian Kroschel of the Karlsruhe Institute of Technology and the Fraunhofer
IOSB and Prof. Andreas Wendemuth of the Otto-von-Guericke-University in
Magdeburg.
Forguidancealongthispathuptothedoctorallevel,Ithankmyformeradvisor
Prof. Manfred K. Lang again at this point.
Next, I would particularly like to thank my precursor Prof. Günther Ruske for
yielding his lectures on Digital Processing of Speech Signals, Automatic Pattern
Recognition in Speech Processing and Data Analysis and Information Reduction
to myself and of course for his countless advice.
Aspecial thankyou isalsodedicatedtomysuperiors atthe CNRS-LIMSI and
itsSpokenLanguageProcessingGroupinOrsay,France,Prof.LaurenceDevillers,
Dr. Jean-Luc Gauvain and his wife Dr. Lori Lamel and Dr. Joseph-Jean Mariani,
as well as to my colleagues at the time, Dr. Matthias Brendel and Dr. Riccardo
Zaccarelli for the common work and to Clement Chastagnol, Agnes Delaborde,
Nicolas Foucault, Marie Tahon and Christophe Vaudable for the manifold
discussions—merci à tous.
ForwelcomingmeattheImperialCollegeLondon’sDepartmentofComputing
in London, UK, and for our constant co-operation at the time of writing, I would
also like to particularly thank Prof. Maja Pantic and her excellent team of the
IntelligentBehaviourUnderstandinggroup(i-bug),inparticularDr.HaticeGunes
xi
xii Acknowledgments
(Queen Mary University of London) and Dr. Michel Valstar (University of Not-
tingham) for our common research.
For hosting me as guest researcher at the National ICT Australia (NICTA) in
Sydney, Australia, and for their infinite hospitality, I would next like to thank
Dr. Julien Epps and Dr. Fang Chen.
For hosting me as guest lecturer and researcher at the Harbin Institute of
TechnologyinHarbin,China,andendlesscareduringthesedays,Iwanttofurther
express my thanks to Prof. Haifeng Li.
Further,forinvitingmeaslecturerinIntelligentAudioAnalysisandhostingme
attheUniversitàPolitecnicadelleMarcheinAncona,Italy,aswellasformanifold
commonworkandorganisationwiththeA3Lab,IwouldliketothankDr.Stefano
Squartini and Prof. Francesco Piazza and from their team in particular
Dr. Emanuele Principi and Dr. Rudy Rotili—grazie mille tutti.
For their outstanding contributions, I would further like to thank my team of
post doctoral and doctoral students at the Institute for Human–Machine Com-
munication, namely Dr. Silvia Monica Feraru, Dr. Cyril Joder, Dr. Fabien Ring-
evalandJunDeng,FlorianEyben,ErikMarchi,FelixWeninger,MartinWöllmer
and Zixing Zhang.
Next, I would like to thank the external doctoral student Raymond Brückner,
and those who visited my working group in ‘sandwich doctoral programs’,
namely, Wenjing Han (Harbin Institute of Technology, China) and Masao
Yamagichi (Tokyo Institute of Technology, Japan).
Also,Iwouldliketothanktheresearchassistantsfortheircontributionsinthis
period in temporal order from the beginning: Thomas Hook, Claudia Tiddia,
Fabian Bross, Erik Marchi, Jordi Feliu, Simone antke, Thomas Knauer, Otto-
ThadäusJandlandYoussefAdel,aswellasthestudenttraineesSunYang,Judith
Köppe and Salma Zouari.
Averywarmandspecialthankyouisalsodedicatedtoallthemaster’sstudents
at the Institute for Human–Machine Communication of this period, who pursued
their theses under my guidance, in temporal order Thomas Mikschl, Ricardo
Minguez, Naijiang Lu, Florian Dibiasi, Hermann Anton Karl, Jin Yao, Martin
Wöllmer, Florian Eyben, Yuan Ma, Josef Zwack, Sebastian Kopp, Xiaodi Wang,
Cong Hui, Peter Grosche, Christoph Scheuermann, Tobias Knaup, Fabian He-
inemann, Yang Sun, Haroon Jacob, Yunpeng Sun, Felix Weninger, Benedikt
Gollan, Xiaohua Zhang, Christian Landsiedel, Erik Marchi and Duc Bui Tran, as
wellasthebachelorstudents—alsointemporalorder—MoritzDausinger,Florian
Eyben, Qianqian Xu, Benedikt Gollan, Xiaohua Zhang, Ugur Özcan, Kayssar
Chaabane, Chi He, Haitao Liu, Jian Li, Clemens Hage, Johannes Dorfner, Jürgen
Glatz, Christoph Kozielski, Judith Köppe and Felix Friedmann, and the interdis-
ciplinary project students Damir Ismailovic, Kevin Tunstall, Maximilian Wöhrl,
ChristianKern,LorenzMösenlechner,TayfurCoskun,AlexanderLehmann,Felix
Weninger and Marcel Knapp.
Among the many colleagues at the Institute in Munich, I will mention only a
few from these years, but thank them all for unforgettable years and exchanges: