Human and Automatic Speaker Recognition over Telecommunication Channels

178 Pages·2016·3.955 MB·English

by  Laura Fernández Gallardo (auth.)

T-Labs Series in Telecommunication Services
Laura Fernández Gallardo
Human and Automatic
Speaker Recognition
over Telecommunication
Channels
T-Labs Series in Telecommunication Services
Series editors
Sebastian Möller, Berlin, Germany
Axel Küpper, Berlin, Germany
Alexander Raake, Berlin, Germany
More information about this series at http://www.springer.com/series/10013
á
Laura Fern ndez Gallardo
Human and Automatic
Speaker Recognition
over Telecommunication
Channels
123
Laura FernándezGallardo
University of Canberra
Canberra,ACT
Australia
ISSN 2192-2810 ISSN 2192-2829 (electronic)
T-Labs Series in Telecommunication Services
ISBN978-981-287-726-0 ISBN978-981-287-727-7 (eBook)
DOI 10.1007/978-981-287-727-7
LibraryofCongressControlNumber:2015946762
SpringerSingaporeHeidelbergNewYorkDordrechtLondon
©SpringerScience+BusinessMediaSingapore2016
Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpart
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
orinformationstorageandretrieval,electronicadaptation,computersoftware,orbysimilarordissimilar
methodologynowknownorhereafterdeveloped.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publicationdoesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfrom
therelevantprotectivelawsandregulationsandthereforefreeforgeneraluse.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authorsortheeditorsgiveawarranty,expressorimplied,withrespecttothematerialcontainedhereinor
foranyerrorsoromissionsthatmayhavebeenmade.
Printedonacid-freepaper
SpringerScience+BusinessMediaSingaporePteLtd.ispartofSpringerScience+BusinessMedia
(www.springer.com)
Preface
The automatic detection of people’s identity from their voices is part of modern
telecommunication services. This generally requires the telephone transmission of
speech to remote servers that perform the recognition task. The transmission may
introduce severe distortions that degrade the system performance and hence rep-
resents one of the major challenges speech technologies are currently facing.
Similarly, humans also cope with the difficulty of reliably identifying talkers from
speechtransmittedovercommunicationchannels,particularlyiftheutteranceheard
is of short duration.
This book addresses the evaluation of the human and of the automatic speaker
recognition performances under different channel distortions caused by bandwidth
limitation, codecs, and electroacoustic user interfaces, among other impairments.
The main contribution of this work is the demonstration of the benefits of com-
munication channels of extended bandwidth, together with an insight into how
speaker-specific characteristics of speech are preserved through different trans-
missions. This book intends to assist students, researchers, and engineers to assess
the speaker recognition performance employing transmitted speech. Particularly
interesting for network engineers, this work provides sufficient motivation for
consideringspeakerrecognitionasacriterionforthemigrationfromnarrowbandto
enhanced bandwidths, such as wideband and super-wideband.
Thisbookwas written inthecontextofmyPh.D. project,a research agreement
between the Telekom Innovation Laboratories and the Technische Universität
Berlin (TU Berlin), Germany, and the University of Canberra (UC), Australia. It
therefore involved periods of research in Berlin, at the Quality and Usability Lab
of the TU Berlin, and in Canberra, at the Human-Centred Computing Laboratory
of the UC.
This work would not have been possible without the contributions of many
supporters. I am most thankful to my two main supervisors Prof. Michael Wagner
and Prof. Sebastian Möller for their support, direction, and advice throughout the
courseofmyPh.D.Theirinsightfulcommentsandfeedbackhaverepeatedlyguided
my research focus towards its final form. My gratitude goes also to my third
v
vi Preface
supervisor Associate Professor Roland Göcke, for his constructive comments and
inputonmywork.IwishtoexpressmyappreciationtotheUniversityofCanberra
fortheresearchgrantawardandtoDeutscheTelekomAGforfullprojectfunding.
Iwouldliketothank,amongmycolleaguesinBerlin,thegroupsofspeakersand
listeners who volunteered to be recorded and to participate in my auditory tests.
This was a kind favour as I required test participants who knew each another and
had been exposed for a long period to one another’s voices. Special thanks to
Marcel Wältermann and to Janto Skowronek for their assistance with the audio
transmissionandreceptionscenariosoftheauditorytestsIconducted.Iwouldalso
like to acknowledge Friedemann Köster’s help in estimating quality of speech
signals employing instrumental speech quality measures.
After having acquired a background on human voice perception and speaker
recognisability, I was still new to automatic speaker recognition procedures.
I would like to thank my colleague David Vandyke for the quick introduction to
speaker recognition methodology when I arrived to Australia, jointly with Prof.
Michael Wagner, which saved me considerable time.
Last, but not least, thanks to my family for their support, understanding, and
patiencewhileIwas17,800kmawayfromhome(thecurrent2,300kmseemstobe
a bearable distance), and thanks to my friends, for their always-comforting words.
Among other incentives, this encouraged my continuous work and dedication
towards the completion of this book.
Berlin, June 2015 Laura Fernández Gallardo
Contents
1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Scope of This Book and Contribution. . . . . . . . . . . . . . . . . . . . 3
1.3 Outline of This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Literature Review. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Today’s Communication Channels and Their Main
Impairments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Channel Quality Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Subjective Speech Quality Assessment . . . . . . . . . . . . . . 9
2.2.2 Instrumental Speech Quality Measures . . . . . . . . . . . . . . 11
2.2.3 Relations Between Quality and Other Attributes
of the Speech Signal. . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Human Speaker Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.1 Speech Characteristics Enabling Human Speaker
Recognition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.2 Effects of Communication Channels on Human
Speaker Recognition. . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.3 Literature on Human Speaker Recognition
and This Book. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Automatic Speaker Recognition . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4.1 Automatic Speaker Recognition Principles
and Main Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.2 Effects of Phonetic Content on Automatic Speaker
Recognition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.3 Effects of Communication Channels on Automatic
Speaker Recognition. . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.4 NIST Speaker Recognition Evaluations. . . . . . . . . . . . . . 31
vii
viii Contents
2.4.5 Comparison Between the Human and the Automatic
Speaker Recognition Performance. . . . . . . . . . . . . . . . . . 32
2.4.6 Literature on Automatic Speaker Recognition
and This Book. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3 Human Speaker Identification Performance Under
Channel Degradations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1 Experimental Environment of the Listening Tests. . . . . . . . . . . . 35
3.1.1 Database Collection for the Listening Tests. . . . . . . . . . . 35
3.1.2 Listening Test 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1.3 Listening Test 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 Effects of Codec and Bandwidth . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Effects of Electro-Acoustic User Interface . . . . . . . . . . . . . . . . . 46
3.4 Effects of Random Packet Loss . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Target Speaker and Familiarity. . . . . . . . . . . . . . . . . . . . . . . . . 48
4 Importance of Intelligible Phonemes for Human Speaker
Recognition in Different Bandwidths. . . . . . . . . . . . . . . . . . . . . . . 51
4.1 Human Speaker Recognition from Logatomes . . . . . . . . . . . . . . 51
4.1.1 Audio Preparation and Listening Test. . . . . . . . . . . . . . . 52
4.1.2 Accuracies per Logatome in Different Bandwidths . . . . . . 53
4.2 Human Speech Intelligibility from Logatomes . . . . . . . . . . . . . . 57
4.3 Relation Between Speaker Recognition and Intelligibility
in Narrowband and in Wideband . . . . . . . . . . . . . . . . . . . . . . . 62
5 Automatic Speaker Verification Performance Under
Channel Distortions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.1 Datasets and Speech Transmissions. . . . . . . . . . . . . . . . . . . . . . 64
5.2 Effects of Channel Impairments in Matched Conditions. . . . . . . . 65
5.2.1 GMM-UBM Performance Under Channel Distortions. . . . 67
5.2.2 JFA Performance Under Bandwidth
and Codec Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2.3 I-Vectors Performance Under Bandwidth
and Codec Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3 Training and Testing Approaches to Reduce Possible
Mismatch Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.3.1 JFA Under Channel Mismatch. . . . . . . . . . . . . . . . . . . . 77
5.3.2 I-Vectors Under Channel Mismatch . . . . . . . . . . . . . . . . 80
6 Detecting Speaker-Discriminative Spectral Content
in Wideband for Automatic Speaker Recognition. . . . . . . . . . . . . . 85
6.1 Effects of the Transmission Channel on the Distribution
of Speaker-Discriminative Spectral Content . . . . . . . . . . . . . . . . 86
6.1.1 Audio Material. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Contents ix
6.1.2 Spectral Sub-band Analysis. . . . . . . . . . . . . . . . . . . . . . 87
6.1.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.1.4 Sub-band Score-Level Fusion. . . . . . . . . . . . . . . . . . . . . 94
6.2 Different Cepstral Features for Narrowband
and for Wideband Speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.2.1 Speech Filtering and Feature Extraction. . . . . . . . . . . . . . 97
6.2.2 I-Vector Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2.4 Score-Level Fusion of Two Frequency Ranges. . . . . . . . . 105
6.3 Relevance of Phonetic Information Under Transmission
Channel Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3.1 Phoneme Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.3.2 I-Vector Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.3.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 109
7 Relations Among Speech Quality, Human Speaker
Identification, and Automatic Speaker Verification . . . . . . . . . . . . 113
7.1 Quality and Performance Metrics for Different
Channel Degradations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.1.1 Instrumental Quality Measurements . . . . . . . . . . . . . . . . 114
7.1.2 Quality and Speech and Speaker Recognition
Performance Metrics. . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2 Predicting Human Speaker Identification Performance
from Measured Speech Quality. . . . . . . . . . . . . . . . . . . . . . . . . 119
7.2.1 Model Fit with POLQA MOS as Estimator. . . . . . . . . . . 119
7.2.2 Model Fit with DIAL Coloration as Estimator . . . . . . . . . 123
7.2.3 Estimations of Human Speaker Identification
Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.3 Predicting Automatic Speaker Verification Performance
from Measured Speech Quality. . . . . . . . . . . . . . . . . . . . . . . . . 131
7.3.1 New Instrumental Quality Measurements. . . . . . . . . . . . . 131
7.3.2 Model Fit with POLQA MOS as Estimator. . . . . . . . . . . 133
7.3.3 Model Fit with DIAL Coloration as Estimator . . . . . . . . . 133
7.3.4 Estimations of Automatic Speaker Verification
Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
7.4 Predicting Human Speaker Identification Performance
from Computed Speaker Verification EERs . . . . . . . . . . . . . . . . 138
7.4.1 Model Fit with EERs as Estimators . . . . . . . . . . . . . . . . 138
7.4.2 Estimations of Human Speaker Identification
Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139