An overview of Automatic Speaker Recognition
Gérard [email protected]
GET-ENST/CNRS-LTCI46 rue Barrault
75634 PARIS cedex 13http://www.tsi.enst.fr/~chollet
Outline
Motivations, Applications Speech production background Speaker characteristics in the speech signal Automatic Speaker Verification :
Decision theory Text dependent Text independent
Databases, Evaluation, Standardization Audio-visual speaker verification Conclusions Perspectives
Why should a computer recognize
who is speaking ?
Protection of individual property (habitation, bank account, personal data, messages, mobile phone, PDA,...)
Limited access (secured areas, data bases) Personalization (only respond to its master’s
voice) Locate a particular person in an audio-visual
document (information retrieval) Who is speaking in a meeting ? Is a suspect the criminal ? (forensic applications)
Domains of Automatic Speaker Recognition
Your voice is a signature Speaker verification (Voice Biometric)
Are you really who you claim to be ? Identification within an open set :
Is this speech segment coming from a known speaker ?
Identification within a closed set Speaker detection, segmentation, indexing,
retrieval : Looking for recordings of a particular speaker
Combining Speech and Speaker Recognition Adaptation to a new speaker Personalization in dialogue systems
Applications
Access Control Physical facilities, Computer networks,
Websites Transaction Authentication
Telephone banking, e-Commerce Speech data Management
Voice messaging, Search engines Law Enforcement
Forensics, Home incarceration
Voice Biometric
Avantages Often the only modality over the telephone, Low cost (microphone, A/D), Ubiquity Possible integration on a smart (SIM) card Natural bimodal fusion : speaking face
Disadvantages Lack of discretion Possibility of imitation and electronic
imposture Lack of robustness to noise, distortion,… Temporal drift
Speaker Identity in Speech
Differences in Vocal tract shapes and muscular control Fundamental frequency (typical values)
100 Hz (Male), 200 Hz (Female), 300 Hz (Child) Glottal waveform Phonotactics Lexical usage, idiolects
The differences between Voices of Twins is a limit case
Voices can also be imitated or disguised
spectral envelope of / i: /
f
A
Speaker A
Speaker B
Speaker Identity
segmental factors (~30ms)
glottal excitation:fundamental frequency, amplitude,voice quality (e.g., breathiness)
vocal tract:characterized by its transfer function and represented by MFCCs (Mel Freq. Cepstral Coef)
suprasegmental factors speaking speed (timing and rhythm of speech units) intonation patterns dialect, accent, pronunciation habits
Speech production
Speech analysis
Inter-speaker Variability
We wereaway
ayear ago.
Intra-speaker Variability
We
were
away
a
year
ago.
Mel Frequency Cepstral Coefficients
Speaker Verification
Typology of approaches (EAGLES Handbook) Text dependent
Public password Private password Customized password Text prompted
Text independent Incremental enrolment Evaluation
Automatic Speaker Verification
Claimed Identity Automatic Speaker Verification System
Acceptation
Rejection
Speech processing Biometric Technology
What are the sources of difficulty ?
Intra-speaker variability of the speech signal (due to stress, pathologies, environmental conditions,…)
Recording conditions (filtering, noise,…) Temporal drift Intentional imposture Voice disguise
Two types of errors : False rejection (a client is rejected) False acceptation (an impostor is accepted)
Decision theory : given an observation O and a claimed identity H0 hypothesis : it comes from an impostor H1 hypothesis : it comes from our client
H1 is chosen if and only if P(H1|O) > P(H0|O) which could be rewritten (using Bayes law) as
Decision theory for identity verification
)1()(
)(
)1(
HPHoP
HoOP
HOP>
)1()(
)(
)1(
HPHoP
HoOP
HOP>
Decision
Distribution of scores
Receiver Operating Characteristic (ROC) curve
Detection Error Tradeoff (DET) Curve
History of Speaker Recognition
Current approaches
Text-dependent Speaker Verification
Uses Automatic Speech Recognition techniques (DTW, HMM, …)
Client model adaptation from speaker independent HMM (‘World’ model)
Synchronous alignment of client and world models for the computation of a score.
Dynamic Time Warping (DTW)
Dynamic Time Warping (DTW)
meilleurchemin
),()Y,X( 2jid yx∑=μ
“Bonjour” locuteur test Y
“Bon
jour
” lo
cute
ur X
“Bonjour” locuteur 1
“Bonjour” locuteur 2
“Bonjour” locuteur n
DODDINGTON 1974, ROSENBERG 1976, FURUI 1981, etc.
Vector Quantization (VQ)
meilleurequant.
),()Y,X( X2
jiCd y∑=μ
Dictionnaire locuteur 1
Dictionnaire locuteur 2
Dictionnaire locuteur n
“Bonjour” locuteur test Y
Dic
tionn
aire
locu
teur
X
SOONG, ROSENBERG 1987
Hidden Markov Models (HMM)
Bestpath
)S(Plog)Y,X(iXjy∑−=μ
“Bonjour” locuteur 1
“Bonjour” locuteur 2
“Bonjour” locuteur n
“Bonjour” locuteur test Y
“Bon
jour
” lo
cute
ur X
ROSENBERG 1990, TSENG 1992
Ergodic HMM
meilleurchemin
)S(Plog)Y,X(iXjy∑−=μ
HMM locuteur 1
HMM locuteur 2
HMM locuteur n
“Bonjour” locuteur test Y
HM
M lo
cute
ur X
PORITZ 1982, SAVIC 1990
Gaussian Mixture Models (GMM)
REYNOLDS 1995
An example of a Text-dependent Speaker Verification System :
The PICASSO project Sequences of digits
Speaker independent HMM of each digit Adaptation of these HMMs to the client voice
(during enrolment and incremental enrolment) EER of less than 1 % can be achieved
Customized password The client chooses his password using some
feedback from the system Deliberate imposture
Deliberate imposture
The impostor has some recordings of the target client voice. He can record the same sentences and align these speech signals with the recordings of the client.
A transformation (Multiple Linear Regression) is computed from these aligned data.
The impostor has heard the target client password. He records that password and applies the
transformation to this recording. The PICASSO reference system with less than 1 %
EER is defeated by this procedure (more than 30 % EER)
Incremental enrolment of customised password
The client chooses his password using some feedback from the system.
The system attempts a phonetic transcription of the password.
Incremental enrolment is achieved on further repetitions of that password
Speaker independent phone HMM are adapted with the client enrolment data.
Synchronous alignment likelihood ratio scoring is performed on access trials.
HMM structure depends on the application
Speaker Verification (text independent)
The ELISA consortium ENST, LIA, IRISA, DDL, Uni-Fribourg, Uni-
Balamand... http://elisa.ddl.ish-lyon.cnrs.fr/
NIST evaluations http://www.nist.gov/speech/tests/spk/index.htm
Gaussian Mixture Models, Graphical models Segmental approaches (ALISP)
Gaussian Mixture Model
Parametric representation of the probability distribution of observations:
Gaussian Mixture Models
8 Gaussians per mixture
GMM speaker modeling
Front-endGMM
MODELING
WORLDGMM
MODEL
Front-end GMM model adaptation
TARGETGMM
MODEL
Baseline GMM method
HYPOTH.TARGET
GMM MOD.
Front-end
WORLDGMM
MODEL
Test Speech
xPxPLog ]
)/()/([
λλ
LLR SCORE
λ
λ
)/( λxP
)/( λxP
Λ =
Support Vector Machines and Speaker Verification
Hybrid GMM-SVM system is proposed
SVM scoring model trained on development data to classify true-target speakers access and impostors access,using new feature representation based on GMMs
Modeling
Scoring
GMM
SVM
SVM principles
X (X)
Inpu
t sp
ace
Feat
ure
spac
e Separating hyperplans H , with the optimal hyperplan Ho
Ho
H
Class(X)
Results
State of the art – research directions (3)
world model, speaker independent, train with all available speaker, using the
algorithm EM . client model,
Obtained as an adaptation of , MAP with a prior distribution MLLR with a transform function Unified approach
(Y)pX
(Y)pX
(Y)pX
Adaptation
Degré de liberté variable Partitionnement variable des distributions Après chaque étape E de l’EM partitionnement
donnant une quantité de données suffisante par classe
12
12
9
17
6
23
21
33
56
Hierarchical - MLLR adapted System
National Institute of Standards & Technology (NIST)
Speaker Verification Evaluations
• Annual evaluation since 1995• Common paradigm for comparing technologies
Evaluations NIST: généralités
Standard reconnu pour l’évaluation des systèmes de vérification du locuteur
Plusieurs centaines de locuteurs différents, Plusieurs dizaines de milliers d’accès de test.
Participation des meilleurs laboratoires mondiaux MIT, IBM, Nuance….
Participation de l’ENST depuis 1997.
Evaluations NIST: Protocole
Phase d’apprentissage 2 minutes de parole spontanée Condition téléphonique, réseau cellulaire
Phase de test Durée des fichiers de 5s à 50s de parole
spontanée
Evaluations NIST: Résultats
Les résultats sont présentés et discutés lors d’un workshop annuel.
Amélioration constante des performances de l’ENST (18%9%) malgré une augmentation de la difficulté: Réduction de la durée d’apprentissage, Réseau commuté réseau cellulaire.
Evaluations NIST: Résultats
ENST 2003
Combining Speech Recognition and Speaker Verification.
Speaker independent phone HMMs Selection of segments or segment classes
which are speaker specific Preliminary evaluations are performed on the
NIST extended data set (one hour of training data per speaker)
1. 1 Speech Segmentation
Large Vocabulary Continuous Speech Recognition (LVCSR) need huge amount of transcribed speech data language (and task) dependent good results for a small set of languages (with existing AND
available transcripts) we do not have such system
Data-driven speech segmentation not yet usable for speech recognition purposes no annotated databases needed language and task independent we could use it to segment the speech data for a
text-independent speaker verification task and for language identification
ALISP (Automatic Language Independent Speech Processing) method
1.2 ALISP data-driven speech segmentation
3. Data-driven Speech Segmentation
for Speaker Verification
Current best speaker verification systems are based on Gaussian Mixture Models (each speech frame is treated independently, and no temporal information is taken into account);
Improvements are still necessary Speech is composed of different sounds Phonemes have different discriminant characteristics for
speaker verification nasals and vowels convey more speaker characteristics
then other speech classes we would like to exploit this idea, but with data-driven
ALISP unit An automatic speech segmentation tool is needed
3.1 Advantages and disadvantages of the speech segmentation step
Problems: Need of an automatic speech segmentation tool Speaker modeling per speech classes => more data
needed More classes => more complicated systems
Advantages Possibility to use it in combination with a dialogue
based systems Text-prompted speaker verification Better accuracy if enough speech data available
3.2 Proposed system: ALISP based Segmental Speaker Verification using
DTW
Speaker specific information is extracted from the : ALISP based speech segments = > Client Dictionary
Non-speaker (world speakers) : ALISP based speech segments => World Dictionary
Dynamic Time Warping (DTW) was already used for speaker verification, but in a text-dependent mode
comparison of two speech data with a similar linguistic content
the DTW distance measure between two speech segments conveys some speaker specific characteristics
Originality: use DTW in text-independent mode The speech data are first segmented in ALISP classes, in
order to remove the linguistic variability Measure the distances among speaker and non-speaker
speech segments
3.3 Searching in client and world speech dictionaries
for speaker verification purposes
3.4 Database and experimental setup for the
speaker verification experiments
Development data: NIST 2001 cellular data (American English)
world speakers (60 female + 59 male): train the ALISP speech segmenter model the non-speakers
Evaluated on small subset (14 female + 14 male speakers) from
NIST 2001 cellular data full set of NIST 2002 cellular data (??? speakers)
Speech parameterization : LPCC for initial ALISP segmentation and MFCC afterward
64 ALISP speech classes
3.5 Results: example of data-driven speech segmentation for speaker verification
Comparison of a manual transcription with the ALISP segmentation (I think my my daughter )
2 occurrences of the English phone-sequence : m - ay ; corresponding ALISP sequences: HM-Hf-Ha and
HM-Hz-Ha-HC
3.6 Results: another example data-driven speech segmentation for speaker verification
2 another occurrences of the English phone : ay ; the corresponding ALISP sequences: HX-Hf and Hf-Ha previous slide : Hf-Ha and Ha-
Hz
3.7 Speaker Verification DET curves
3.8 Conclusions
State of the art NIST 2002 results for EER: best 8% to worst 28%
Problem with the small data set results: influence of the size of the test set and/or mismatched train/test conditions
What we have NOT done: exploit the speech classes (silence classes are also
included) normalization (with pseudo-impostors) exploit the DTW distance value, not only the
“preference” result
SuperSID experiments
GMM with cepstral features
Selection of nasals in words in -ing
being everythi
ng getting
anything thing
something
things going
Fusion
Fusion results
Visages parlants et vérification d’identité
Le visage et la parole offrent des informations complémentaires sur l’identité de la personne.
De nombreux PC, PDA et téléphones sont et seront équipés d’une caméra et d’un microphone
Les situations d’imposture sont plus difficiles à réaliser.
Thème de recherche développé à l’ENST dans le cadre du projet IST-SecurePhone
Visages parlants et vérification d’identité
Série de chiffres (PIN code) Mot de passe personnalisé
Fusion Parole et Visage
(thèse de Conrad Sanderson, août 2002)
1. Acquisition des signaux biométriques pour chaque modalité2. Calcul du score de décision pour chaque système3. Calcul d’un score de décision final basé sur la fusion des scores
mono-modalité
InsecureNetwork
Serveur distant:1. Accès à des services sécurisés2. Validation de transactions3. Etc.
Exemple d’application
Conclusions et Perspectives La parole permet une vérification d’identité
à travers le téléphone.
Combiner les approches dépendantes et indépendantes du texte améliore la fiabilité.
Si l’on utilise le visage pour vérifier l’identité, il ne coûte pas cher d’ajouter la parole (et cela rapporte gros !).
De plus en plus de PC, PDA et téléphones sont équipés d’un microphone et d’une caméra. La reconnaissance audio-visuelle devrait se généraliser.
Perspectives
Speech is often the only usable biometric modality (over the telephone network).
Fusion of modalities.
A number of R&D projects within the EU.