16/05/2003Reunion Bayestic / Murat Deviren1 Reunion Bayestic Excuse moi! Murat Deviren.

Slides:



Advertisements
Présentations similaires
Les pronoms compléments
Advertisements


4 Avoir Les normes: –Communication 1.2 : Understanding the written and spoken language –Comparisons 4.1 : Understanding the language through making comparisons.
Département fédéral de lintérieur DFI Office fédéral de la statistique OFS Implementing the economic classification revision (NACE / ISIC) in the Business.
Practical Session – Defining Learning Outcomes
THALES Communications Les informations contenues dans ce document sont la propriété exclusive du Groupe THALES. Elles ne doivent pas être divulguées sans.
Apprentissage semi-supervisé
(Nom du fichier) - D1 - 01/03/2000 FTR&D/VERIMAG TAXYS : a tool for the Development and Verification of RT Systems a joint project between France Telecom.
Le Passé Composé J'ai fini Elle a dansé Il a voyagé
1 La bibliométrie pour l'évaluation stratégique des institutions de recherche : usages et limites Indicators for strategic positioning of the research.
Gérard CHOLLET Fusion Gérard CHOLLET GET-ENST/CNRS-LTCI 46 rue Barrault PARIS cedex 13
Talking about your House How to say where things are, what they are like, and to whom they belong.
Inforoute Santé du Canada Les défis de linteropérabilité en e-santé Mike Sheridan, Chef de lexploitation 19 mai 2006.
Talking about yourself
interaction in the .LRN platform
WHAT/WHICH ONE? THIS ONE!. Review the forms of quel meaning what or which: MasculineFeminine Singular Plural.
Le Superlatif des Adjectifs
Status report SOLEIL April 2008
Formal/Theory Phenomenology/Ex periments chaos break-up, giant-resonances, fusion interdisciplinarity (clusters, bose) mean-field (as a general theory)
Delagnes 15/10/07 1 Resist Meeting Saclay 15/10/07 E. Delagnes.
1 Découverte des Outils SI de Cadence Ecole dElectronique Numérique IN2P3 Roscoff 2006 Découverte des Outils dAnalyse dIntégrité du Signal de Cadence ®
TP2 ... MVC ? JList JLabel JSlider ImageLibrary Contrôleur Vue Modèle
Minimisation Techniques 1 Assimilation Algorithms: Minimisation Techniques Yannick Trémolet ECMWF Data Assimilation Training Course March 2006.
Université Des Sciences Et De La Technologie DOran Mohamed Boudiaf USTO République Algérienne Démocratique et Populaire Département de linformatique Projet.
Defence R&D Canada R et D pour la défense Canada Novel Concepts for the COP of the Future Denis Gouin Alexandre Bergeron-Guyard DRDC Valcartier.
Y and en Two little words with a lot of meaning. y.
Bayesian Inference Algorithms Revisited
Rules Each group answers every question. A student will be selected at random to answer the question. If that team misses, a 2 nd name is drawn, and they.
TM.
Une Amie Un Ami Français I.
____________________ Pourquoi? L/O: To be able to justify your opinion about school subjects STARTER: Trouve les paires Match the French to the English:
Quest-ce que tu fais à lécole? Buts: Using regular verbs to say what we do at school.
Defence Research and Development Canada Recherche et développement pour la défense Canada Canada 11-1.
Assessment and the new secondary curriculum S. Barfoot.
How to solve biological problems with math Mars 2012.
Discussion, Youth Engagement, and Appreciation of Diversity Kelly Campbell 1, Linda Rose-Krasnor 1, Michael Busseri 1, Mark Pancer 2 and the Centre of.
AFNOR NF Z – "Online Consumer Reviews
Indefinite articles, plural of nouns
TortoiseSVN N°. Subversion : pour quoi faire ? Avoir un espace de stockage commun – Tous les étudiants du SIGLIS ont un espace svn commun Partager vos.
PURCHASING PHASE REVIEW Cornerstones of Purchase baseline
Les choses que j aime Learning Objective: To know how to use j aime to talk about things I like to do.
Laboratoire de Bioinformatique des Génomes et des Réseaux Université Libre de Bruxelles, Belgique Introduction Statistics.
L’ensemble microcanonique
1.
ETL et Data Mining Présenté par : Marc Catudal-Gosselin Université de Sherbrooke automne 2004 automne 2004.
Présentation dun modèle dinterface adaptative dun système de diagnostique et dintervention industriel: ADAPTS (Adaptive Diagnostics And Personalized Technical.
Passage entre quaternions et matrice des cosinus directeurs Transition from Quaternions to Direction Cosine Matrices.
Marketing électronique Cours 5 La personnalisation.
Les normes: Communication 1.2 Comparisons 4.2 La question essentielle: What is the formula for conjugating -RE verbs and what are some of these verbs?
Thematic Alignment of Static Documents with Meeting Dialogs Dalila Mekhaldi Diva Group Department of Computer Science University of Fribourg.
J. Duchesne1, P. Raimbault2 and C. Fleurant1
Dominique Vaufreydaz, ESSLLI ASR and scalability Dominique Vaufreydaz ESSLLI02.
Integer Caratheodory theorems. Linear Caratheodory Given A={a 1,…, a n } IR d. For all v cone(A) there exists B A, |B| d st v cone(B) Proof : Si A nest.
Français I. Une fille française Gabrielle est française. Elle est blonde. Elle est belle. Elle est de Paris.
La prononciation française
INDICATOR DEFINITION An indicator describes the manifestation of a process of change resulting from the pursuit of an action. Un indicateur décrit la manifestation.
6 Le verbe Faire Les normes: –Communications 1.2: Understanding the written and spoken language –Comparisons 4.1: Understanding language through comparisons.
Branche Développement Le présent document contient des informations qui sont la propriété de France Télécom. L'acceptation de ce document par son destinataire.
VTHD PROJECT (Very High Broadband Network Service): French NGI initiative C. GUILLEMOT FT / BD / FTR&D / RTA
KM-Master Course, 2004 Module: Communautés virtuelles, Agents intelligents C3: Collaborative Knowledge construction & knowledge sharing Thierry NABETH.
Formatting of Game © Candace R. Black, Al rights reserved. Révisions!
When do we use numbers? Why are they important? Why is it important to know numbers in French (or any other language)? Can you think of some REAL WORLD.
Ministère de l’Éducation, du Loisir et du Sport Responsables des programmes FLS et ELA: Diane Alain et Michele Luchs Animateurs: Diane Alain et Michael.
Information Theory and Radar Waveform Design Mark R. bell September 1993 Sofia FENNI.
Le genre musical François Pachet. Sur Amazon ? Pop General Adult Contemporary Britpop Dance Pop Disco Easy Listening Emerging Artists Latin Pop Motown.
I can use longer and more complex sentences by understanding and using comparisons.
1 Linear Prediction. 2 Linear Prediction (Introduction) : The object of linear prediction is to estimate the output sequence from a linear combination.
IP Multicast Text available on
Transcription de la présentation:

16/05/2003Reunion Bayestic / Murat Deviren1 Reunion Bayestic Excuse moi! Murat Deviren

16/05/2003Reunion Bayestic / Murat Deviren2 Contents Frequency and wavelet filtering Supervised-predictive compensation Language modeling with DBNs Hidden Markov Trees for acoustic modeling

16/05/2003Reunion Bayestic / Murat Deviren3 Contents Frequency and wavelet filtering Supervised-predictive compensation Language modeling with DBNs Hidden Markov Trees for acoustic modeling

16/05/2003Reunion Bayestic / Murat Deviren4 Frequency Filtering Proposed by Nadeu95, Paliwal99. Goal : Spectral features comparable with MFCCs Properties : –Quasi decorrelation of logFBEs. –Cepstral weighting effect –Emphasis on spectral variations FF1FF2FF3 H(z)1-z -1 z-z -1 1-z -2 logFBEs DCT H(z) MFCC FF H(z) = 1-az -1 Simplified block diagram for MFCC and FF parameterizations Typical derivative type frequency filters

16/05/2003Reunion Bayestic / Murat Deviren5 Evaluation of FF on Aurora-3 Significant performance decrease for FF2 & FF3 in high mismatch case FF1FF2FF3 H(z)1-z -1 z-z -1 1-z -2

16/05/2003Reunion Bayestic / Murat Deviren6 Wavelets and Frequency Filtering FF1 = Haar Wavelet Reformulate FF as wavelet filtering Use higher order Daubechies wavelets Promising results Published in ICANN 2003

16/05/2003Reunion Bayestic / Murat Deviren7 Perspectives BUT –These results could not be verified on other subsets of Aurora-3 database. To Do –Detailed analysis of FF and wavelet filtering –Develop models that exploit frequency localized features. –Exploit statistical properties of wavelet transform.

16/05/2003Reunion Bayestic / Murat Deviren8 Contents Frequency and wavelet filtering Supervised-predictive compensation Language modeling with DBNs Hidden Markov Trees for acoustic modeling

16/05/2003Reunion Bayestic / Murat Deviren9 Noise Robustness Signal processing techniques : –CMN, RASTA, enhancement techniques Compensation schemes –Adaptive : MLLR, MAP Requires adaptation data and a canonical model –Predictive : PMC Hypothetical errors in mismatch function Strong dependence on front-end parameterization Multi-condition training

16/05/2003Reunion Bayestic / Murat Deviren10 Supervised-predictive compensation Goal : –exploit available data to devise a tool for robustness. Available data : –speech databases recorded in different acoustic environments. Principles : –Train matched models for each condition. –Train noise models. –Construct a parametric model that describe how matched models vary with noise model.

16/05/2003Reunion Bayestic / Murat Deviren11 Supervised-predictive compensation Advantages : –No mismatch function –Independent of front-end –Canonical model is not required –Computationally efficient –Model can be trained incrementally i.e. can be updated with new databases

16/05/2003Reunion Bayestic / Murat Deviren12 Deterministic model Databases : D 1, …, D K Noise conditions : n 1, …, n K S w (k) : matched speech model for acoustic unit w W trained on noise condition n k. N {1,…, K}: noise variable. For each w W, there exists a parametric function f w such that –|| S w (k) – f w (N) || 0 for some given norm ||.||

16/05/2003Reunion Bayestic / Murat Deviren13 Probabilistic model Given –S : speech model parameterization –N : noise model parameterization Learn the joint probability density P(S, N) Given the noise model N, what is the best set of speech models to use? –S` = argmax P(S|N) S1S1 S2S2 S3S3 N1N1 N2N2 N3N3 N S P(S,N) as a static Bayesian network

16/05/2003Reunion Bayestic / Murat Deviren14 A simple linear model Speech model : mixture density HMM Noise model : single Gaussian wls (n k ) = A wls nk + B wls – wls (n k ) : mean vector for mixture component l of state s – nk : mean vector of noise model f w is parameterized with A wls, B wls Supervised training using MMSE minimization

16/05/2003Reunion Bayestic / Murat Deviren15 Experiments Connected digit recognition on TiDigits 15 different noise sources from NOISEX –volvo, destroyer engine, buccaneer…. Evaluations : –Model performance in training conditions –Robustness comparison with multi-condition training : under new SNR conditions, under new noise types.

16/05/2003Reunion Bayestic / Murat Deviren16 Results Even a simple linear model can almost recover matched model performances. The proposed technique can generalize to new SNR conditions and new noise types. Results submitted to EUROSPEECH 2003

16/05/2003Reunion Bayestic / Murat Deviren17 Contents Frequency and wavelet filtering Supervised-predictive compensation Language modeling with DBNs Hidden Markov Trees for acoustic modeling

16/05/2003Reunion Bayestic / Murat Deviren18 Classical n-grams Word probability based on word history. P(W) = i P(w i | w i-1, w i-2, …, w i-n ) w i-n w i-2 wiwi w i-1

16/05/2003Reunion Bayestic / Murat Deviren19 Class based n-grams Class based word probability for a given class history. P(W) = i P(w i | c i ) P(c i | c i-1, c i-2, …, c i-n ) c i-n c i-2 cici c i-1 w i-n w i-2 wiwi w i-1

16/05/2003Reunion Bayestic / Murat Deviren20 Class based LM with DBNs Class based word probability in a given class context. P(W) = i P(w i | c i-n, …, c i,…c i+n ) P(c i | c i-1, c i-2, …, c i-n ) c i-n c i-2 cici c i-1 w i-n w i-2 wiwi w i-1 c i+1 c i+2

16/05/2003Reunion Bayestic / Murat Deviren21 Initial results Training corpus 11 months from le monde ~ 20 million words Test corpus ~ 1.5 million words Vocabulary size : 500 # class labels = 198 wiwi w i-1 cici c i-1 wiwi cici wiwi cici wiwi c i+1 ModelPerplexity

16/05/2003Reunion Bayestic / Murat Deviren22 Perspectives Initial results are promising. To Do –Learning structure with appropriate scoring metric, i.e., based on perplexity –Appropriate back-off schemes –Efficient CPT representations for computational constraints, i.e., noisy-OR gates.

16/05/2003Reunion Bayestic / Murat Deviren23 Contents Frequency and wavelet filtering Supervised-predictive compensation Language modeling with DBNs Hidden Markov Trees for acoustic modeling

16/05/2003Reunion Bayestic / Murat Deviren24 Reconnaissance de la parole à laide de modèles de Markov cachés sur des arbres dondelettes Sanaa GHOUZALI DESA Infotelecom Université Med V - RABAT

16/05/2003Reunion Bayestic / Murat Deviren25 Problèmes de la reconnaissance de la parole Paramétrisation: Besoin de localiser les paramètres du signal parole dans le domaine temps-fréquence Avoir des performances aussi bonnes que les MFCC Modélisation: Besoin de construire des modèles statistiques robuste au bruit Besoin de modéliser les dynamiques fréquentielles du signal parole aussi bien que les dynamiques temporelles

16/05/2003Reunion Bayestic / Murat Deviren26 Paramètrisation La transformée Ondelette a de nombreuses propriétés intéressantes qui permettent une analyse plus fine que la transformée Fourrier; Localité Multi-résolution Compression Clustering Persistence

16/05/2003Reunion Bayestic / Murat Deviren27 Modélisation Il existe plusieurs types de modèles statistiques qui tiennent compte des propriétés de la transformée ondelette; Independent Mixtures (IM): traite chaque coefficient indépendamment des autres (pptés primaire) Markov chains: considère seulement les corrélations entre les coefficients dans le temps (clustering) Hidden Markov Trees (HMT): considère les corrélations entre échelles (persistence)

16/05/2003Reunion Bayestic / Murat Deviren28 Les modèles statistiques pour la transformée ondelette t f t f

16/05/2003Reunion Bayestic / Murat Deviren29 Description du modèle choisi le modèle choisi WHMT : illustre bien les propriété clustering et persistance de la transformée ondelette interprète les dépendances complexes entre les coefficients d'ondelette la modélisation pour la transformée ondelette sera faite en deux étapes: modéliser chaque coefficient individuellement par un modèle de mélange de gaussienne capturer les dépendances entre ces coefficients par le biais du modèle HMT

16/05/2003Reunion Bayestic / Murat Deviren30 Références M. S. Crouse, R. D. Nowak, and R. G. Baraniuk, Wavelet-Based Statistical Signal- Processing Using Hidden Markov Models, IEEE Trans. Signal. Proc., vol. 46, no. 4, pp , Apr M. Crouse, H. Choi and R. Baraniuk, Multiscale Statistical Image Processing Using Tree-Structured Probability Models, IT Workshop, Feb K. Keller, S. Ben-Yacoub, and C. Mokbel, Combining Wavelet-Domain Hidden Markov Trees With Hidden Markov Models, IDIAP-RR 99-14, Aug M. Jaber Borran and R. D. Nowak, Wavelet-Based Denoising Using Hidden Markov Models