Doctoral and Post-doctoral Positions in Signal Processing for Hearing Instruments

Position Description

The ITN ICanHear, starting on 1 January 2013, will provide cutting-edge research projects for 12 doctoral and 5 post-doctoral research fellows in digital signal processing for hearing instruments. ICanHear aims to develop models based on emerging knowledge about higher-level processing within the auditory pathway and exploit that knowledge to develop creative solutions that will improve the performance of hearing instruments.

Attractive grants and a wide variety of international training activities, including collaborations with ICanHear Associated Partners in the U.K., Switzerland, Belgium, U.S.A., and Canada, will be made available to successful candidates, who will stay in the network for a period of 12 to 36 months.

Research and training positions will be available in the following ICanHear labs:

• Institute of Communication Acoustics, Ruhr-Universität Bochum (DE)

• Experimental Oto-rhino-laryngology (ExpORL), Katholieke Universiteit Leuven (BE)

• Institute of Sound and Vibration Research, University of Southampton (UK)

• Hearing Systems Group, Denmark Technical University, Lyngby (DK)

• Laboratory for Experimental Audiology, University Hospital Zurich (CH)

• Signal Processing Group, Siemens Hearing Instruments (DE)

• Cochlear Research and Development, Cochlear Ltd., Mechelen (BE)

Requirements for Candidates and Application procedure:

Early-stage (doctoral) Research Fellows have less than four years experience in research after obtaining a Masters degree in engineering, computer science, or similar.

Experienced (post-doctoral) Researcher Fellows are already in possession of a doctoral degree or have at least 4 years but less than 5 years of research experience in engineering and/or hearing research.

In order to ensure transnational mobility candidates may have resided no more than 12 months (during the last 3 years) in the country of the host institution they wish to apply to. For all positions excellent English language skills are required.

To apply please send in the following documents via e-mail to the ICanHear coordination office (icanhear@rub.de): CV, certified copies of all relevant diplomas and transcripts, two letters of recommendation, proof of proficiency in English, letter of motivation (research interest, reasons for applying to programme and host). For further information on research projects available, application details and eligibility please visit the ICanHear web-site (http://www.icanhear-itn.eu) or contact the project coordinator Rainer Martin (rainer.martin@rub.de).

Back

Top

CAMPAGNE NOEMI HIVER 2012-2013 PROFIL DE POSTE Description de l'Unité Code unité : UMR 7309 Nom de l’unité : Laboratoire Parole et Langage Directeur : Noël NGUYEN Ville : Aix-en-Provence Délégation régionale : DR12 Institut : INSHS Description du poste NUMERO NOEMI : T54030 CORPS : Technicien BAP : C Emploi-type : C4B21-Technicien en instrumentation scientifique, expérimentation et mesure Fonction Technicien de Plateforme Technicien de Plateforme Mission Au sein du Laboratoire Parole et Langage (LPL), affecté au Centre d’Expérimentation sur la Parole (CEP), l’agent sera chargé du soutien aux expériences en collaboration avec le coordinateur de la plateforme. Activités L’activité principale consiste à apporter un soutien quotidien au fonctionnement de la plateforme, il peut s’agir notamment : - D’assurer le prêt et le suivi du matériel utilisé sur la plateforme ou à l’extérieur, - D’effectuer le montage, l’assemblage de sous-ensembles (notamment audio et vidéo) pour la réalisation d’expériences, - D’assister les expérimentateurs lors de la passation d’expériences en appliquant un protocole défini, - D’effectuer des modifications ou adaptations de dispositifs expérimentaux, - D’assurer la maintenance et les interventions de premier niveau, la détection et le diagnostique de pannes,. - De réaliser des enregistrements (installation et enregistrement proprement-dit) audio et vidéo, - D’assurer la gestion des consommables nécessaires au déroulement des expériences, - D’utiliser des applications logicielles de contrôle d’instruments. Compétences Le (ou)la candidate devra faire preuve d’une grande motivation pour ce poste de soutien indispensable au fonctionnement du centre d’expérimentation sur la parole. Une formation de base en électronique et/ou en mesures physiques est souhaitée, pour mener à bien la réalisation éventuelle de modules élémentaires de synchronisation entre instruments, ou encore pour synchroniser les systèmes d’enregistrement audio et vidéo, pour effectuer le montage et l’assemblage de sous-ensembles pour la réalisation de dispositifs expérimentaux. Le ou la candidate devra apprécier le travail en équipe puisqu’il ou elle travaillera en lien étroit avec le coordinateur de la plateforme. La personne doit être capable d’apprendre de nouvelles techniques, et avoir goût pour lele sens du contact humain car elle sera en contact avec un grand nombre d’utilisateurs. Elle devra montrer une grande rigueur dans le respect des procédures mises en place. L’adhésion aux règles d’hygiène et sécurité en place ainsi est indispensable. Contexte Le Laboratoire Parole et Langage est une unité de recherche du CNRS et d’Aix Marseille Université. Il accueille des phonéticiens, linguistes, des informaticiens, des psychologues, des neuroscientifiques, des physiciens et des médecins. Les activités du LPL portent sur l’étude des mécanismes de production et de perception du langage et de la parole. Le LPL se distingue par ses méthodes de recherche reposant à la fois sur l’expérimentation, l’investigation instrumentale et la formalisation. Approche originale dans ce champ scientifique, qui émarge à la fois aux domaines des sciences humaines, des sciences du vivant et des sciences pour l’ingénieur. Cette particularité explique, au-delà d’une forte activité de recherche fondamentale, l’importance des applications développées à partir des travaux menés dans les domaines du traitement de l’écrit, de l’intelligibilité du message parlé, de la conversion texte-parole de qualité, ou encore de l’évaluation et de la rééducation des troubles de la voix ou du langage. Ces caractéristiques font du Laboratoire Parole et Langage une unité de recherche adaptée aux défis scientifiques des sciences du langage, tout en étant impliquée dans leurs enjeux technologiques. Le LPL regroupe actuellement plus de 80 personnes statutaires (chercheurs, enseignants-chercheurs, ingénieurs, techniciens, administratifs), auxquelles s’ajoutent 40 doctorants dont 20 boursiers. Il est le laboratoire français le plus important dans ce domaine scientifique et l’un des premiers en Europe. Le LPL dispose désormais d’une plateforme technique regroupant un ensemble d’instruments pour l’investigation de la production et la perception de la parole : électro-encéphalographie, tracking oculaire, articulographie, électro- palatographie, évaluation articulatoire, etc. Cette ressource unique en Europe est mutualisée au sein du Centre d’Expérimentation sur la Parole (http://www.lpl.univ-aix.fr/~cep/), plateforme technique à laquelle le poste sera affecté le poste. 

Back

Top

Computer Science Internship

CORDIAL group

Title :

Voice Conversion from non-parallel corpora

Description :

The main goal of a voice conversion system (VCS) is to transform the speech

signal uttered by speaker (the source speaker) so that it sounds like it was uttered by an other

person (the target speaker). The applications of such techniques are limitless. For example, a

VCS can be combined to a Text-To-Speech system in order to produce multiple high quality

synthetic voices. In the entertainment domain, a VCS can be used to dub an actor with its own

voice.

State of the art VCS use Gaussian Mixture Models (GMM) to capture the transformation

from the acoustic space of the source to the acoustic space of the target. Most of the models are

source-target joint models that are trained on paired source-target observations. Those paired

observations are often gathered from parallel corpora, that is speech signals resulting from the

two speakers uttering the same set of sentences. Parallel corpora are hard to come with. Moreo-

ver, they do not guaranty that the pairing of vectors is accurate. Indeed, the pairing process is

unsupervised and uses a Dynamic Time Warping under the strong (and unrealistic) hypothesis

that the two speakers truly uttered the same sentence, with the same speaking style. This asser-

tion is often wrong and results in non-discriminant models that tends to over-smooth speaker's

distinctive characteristics.

The goal of this Master subject is to suppress the use of parallel corpora in the process of

training joint GMM for voice conversion. We suggest to pair speech segments on high level speech

descriptors as those used in Unit Selection Text-To-Speech. Those descriptors not only contain

the segmental information (acoustic class for example) but also supra-segmental informations

such as phoneme context, speed, prosody, power, ... In a rst step, both source and target

corpora are segmented and tagged with descriptors. In a second step, each class from one corpus

is paired with the equivalent class from the other corpus. Finally, a classical DTW algorithm

can be applied on each paired class. The expected result is to derive transform models that both

could take into account speaker variability and be more robust to pairing errors.

Keywords :

Voice Conversion, Gaussian Mixture Models

Contacts :

Vincent Barreaud (vincent.barreaud@irisa.fr)

Bibliographie :

[1] H. Benisty and D. Malah. Voice conversion using gmm with enhanced global variance. In

Conference of the International Speech Communication Association (Interspeech)

, pages 669{

672, 2011.

[2] L. Mesbahi, V. Barreaud, and O. Boeard. Non-parallel hierarchical training for voice conver-

sion. In

Proceedings of the 16th European Signal Processing Conference, Lausanne, Switzerland, 2008.

[3] Y. Stylianou, O. Cappe, and E. Moulines. Continuous probabilistic transform for voice conver-

sion.

IEEE Transactions on Speech and Audio Processing, 6(2) :131-142, 1998.

Back

Top

Computer Science Internship

CORDIAL group

Title :

Unit-selection speech synthesis guided by a stochastic model of spectral and prosodic

parameters.

A Text-To-Speech system (TTS) produces a speech signal corresponding to the vocalization

of a given text. Such a system is composed of a linguistic processing stage followed by an acoustic

one which complies as much as possible with the linguistic directives. Concerning the second step,

the most used approaches are

{ the corpus based synthesis approach which lies on the selection and concatenation of unit

sequences extracted from a large continuous speech corpus. It has been popular for 20

years, yielding an unmatched sound quality but still bearing some artefacts due to spectral

discontinuities.

{ the statistical approach. The new generation of TTS systems has emerged in the last years,

reintroducing the rule based systems. The rules are no longer deterministic like in the

rst systems in the 1950's, but they are replaced by stochastic models. HTS, an HMMbased

speech synthesis system, is currently the most used statistical system. The HTS type

systems yield a good acoustic continuum but with a sound quality strongly depending on

the underlying acoustic model.

Recently, some hybrid synthesis systems have been proposed, combining the statistical approach

with the method of unit selection. It consists in using the acoustic descriptions and the

melodic contours generated by a statistical system in order to drive the cost function during the

natural speech unit selection phase, or also, substituting the poor quality natural speech units

by units derived from a statistical system.

The framework of this subject is the corpus based TTS. Considering the combinatorial problem

due to the search of an optimal unit sequence with a blind sequencing, the work consists

in determining heuristics to reduce the search space and satisfy a real time objective. These

assumptions, based on spectral and prosodic type parameters generated by HTS, will permit to

implement pre-selection lters or to propose new cost functions within the corpus based system

developped by the Cordial group. The production of the hybrid system will be evaluated and

compared via listening tests with standard systems like HTS and a corpus based system.

Keywords :

TTS, Corpus based speech synthesis, Statistical Learning, Experiments.

Contacts :

Olivier Boe
ard, Nelly Barbot, Damien Lolive (prenom.nom@irisa.fr)

Bibliography :

[1] A. W. Black and K. A. Lenzo,

Optimal data selection for unit selection synthesis, 4th ISCA

Tutorial and Research Workshop on Speech Synthesis, 2001.

[2] H. Kawai, T. Toda, J. Ni, M. Tsuzaki and K. Tokuda,

Ximera : a new tts from atr based on

corpus-based technologies

. ISCA Tutorial and Research Workshop on Speech Synthesis, 2004.

[3] S. Rouibia and O. Rosec,

Unit selection for speech synthesis based on a new acoustic target

cost

, Interspeech, 2005.

[4] H. Zen, K. Tokuda and A. W. Black,

Statistical parametric speech synthesis. Speech Communication,

v.51, n.11, pages 1039{1064, 2009.

[5] H. Silen, E. Helander, J. Nurminen, K. Koppinen and M. Gabbouj,

Using Robust Viterbi

Algorithm and HMM-Modeling in Unit Selection TTS to Replace Units of Poor Quality

,

Interspeech 2010.

Back

Top

 Computer Science Internship

CORDIAL group

Title: Grapheme-to-phoneme conversion adaptation using conditional random elds

Description:

Grapheme-to-phoneme conversion consists in generating possible pronuncia-

tions for an isolated word or for a sequence of words. More formally, this conversion is a translit-

eration of a sequence of graphemes, i.e., letters, into a sequence of phonemes, symbolic units to

represent elementary sounds of a language. Grapheme-to-phoneme converters are used in speech

processing



either to help automatic speech recognition systems to decode words from a speech signal



or as a mean to explain speech synthesizers how a written input should be acoustically

produced.

A problem with such tools is that they are trained on large and varied amounts of aligned

sequences of graphemes and phonemes, leading to generic manners of pronouncing words in a

given language. As a consequence, they are not adequate as soon as one wants to recognize

or synthesize specic voices, for instance, accentuated speech, stressed speech, dictating voices

versus chatting voices,

etc. [1].

While multiple methods have been proposed for grapheme-to-phoneme conversion [2, 3], the

primary goal of this internship is to propose a method to adapt grapheme-to-phoneme models

which can easily be adapted under conditions specied by the user. More precisely, the use of

conditional random elds (CRF) will be studied to model the generic French pronunciation and

variants of it [4]. CRFs are state-of-the-art statistical tools widely used for labelling problems

in natural language processing [5]. A further important goal is to be able to automatically

characterize pronunciation distinctive features of a given specic voice as compared to a generic

voice. This means highlighting and generalizing di
erences that can be observed between two

sequences of phonemes derived from a same sequence of graphemes.

Results of this internship would be integrated into the speech synthesis platform of the team

in order to easily and automatically simulate and imitate specic voices.

Technical skills:

C/C++ and a scripting language (e.g., Perl or Python)

Keywords:

Natural language processing, speech processing, machine learning, statistical learn-

ing

Contact:

Gwenole Lecorve (gwenole.lecorve@irisa.fr)

References:

[1] B. Hutchinson and J. Droppo. Learning non-parametric models of pronunciation. In

Pro-

ceedings of ICASSP

, 2011.

[2] M. Bisani and H. Ney. Joint-sequence models for grapheme-to-phoneme conversion. In

Speech

Communication

, 2008.

[3] S. Hahn, P. Lehnen, and Ney H. Powerful extensions to crfs for grapheme to phoneme

conversion. In

Proceedings of ICASSP, 2011.

[4] Irina Illina, Dominique Fohr, and Denis Jouvet. Multiple pronunciation generation using

grapheme-to-phoneme conversion based on conditional random elds. In

Proceedings of

SPECOM

, 2011.

[5] John D. La
erty, Andrew McCallum, and Fernando C. N. Pereira. Conditional random

elds: probabilistic models for segmenting and labeling sequence data. In

Proceedings of

ICML

, 2001.

Back

Top

Post Specification

Post Title:

Ph.D. Researcher in Speech Synthesis

Post Status:

3 years

Department/Faculty:

Centre for Language and Communication Studies (CLCS)

Location:

Phonetics and Speech Laboratory

Salary:

€16,000 per annum (plus fees paid)

Closing Date:

31st January 2013

Post Summary

A Ph.D. Researcher is required to work in the area of speech synthesis at the Phonetics and

Speech Laboratory, School of Linguistic, Speech and Communication Sciences. The position

will involve carrying out research on the topic of Hidden Markov Model (HMM)-based speech

synthesis. Specifically, we are looking for a researcher to work on developing a source-filter

based acoustic modelling for HMM-based speech synthesis which is closely related to the

human speech production process and which can facilitate modification of voice source and

vocal tract filter components at synthesis time.

Background to the Post

Much of the research carried out to date in the Phonetics and Speech Laboratory has been

concerned with the role of the voice source in speech. This research involves the development of accurate voice source processing both as a window on human speech production and for exploitation in voice-sensitive technology, particularly synthesis. The laboratory team is interdisciplinary and includes engineers, linguists, phoneticians and technologists.

This post will the main be funded by the on-going Abair project which has developed the first

speech synthesisers for Irish (www..abair.ie), and the researcher will exploit the current Abair

synthesis platform. In this project the aim is to deliver multi-dialect synthesis with multiple

personages and voices that can be made appropriate to different contexts of use. The post will also be linked to the FastNet project which aims at voice-sensitive speech technologies.

A specific goal of our laboratory team is to leverage our expertise on the voice by improving the naturalness of parametric speech synthesis, as well as making more flexible synthesis platforms which can allow modifications of voice characteristics (e.g., for creating different personalities/characters, different forms of expression etc).

Standard duties of the Post

Initially the researcher will be required to attend some lectures as part of the Masters

programme on Speech and Language Processing. This and a supervised reading

programme will provide a background in the area of voice production, analysis and

synthesis.

* In the very early stages the researcher will be required to develop synthetic voices, using

the Irish corpora, with the standard HMM-based synthesis platform (i.e. HTS). Note that

to work with the Irish corpora does not require a background in the Irish language, as

there will be collaboration with experts in this field.

* The researcher will be required to familiarise themselves with existing speech synthesis

platforms which provide explicit modelling of the voice source (e.g., Cabral et al. 2011,

Raitio et al. 2011, Anumanchipalli et al. 2010).

* The researcher will then need to first implement similar versions of these systems and

then work towards developing novel vocoding methods which would allow full parametric

flexibility of both voice source and vocal tract filter components at synthesis time.

Person Specification

Qualifications

* Bachelors degree in Electrical Engineering, Computer Science with specialisation in

speech signal processing, or related areas.

* Knowledge & Experience (Essential & Desirable)

* Strong digital signal processing skills (Essential)

*Good knowledge of HTS including previous experience developing synthetic voices

(Essential)

* Knowledge of speech production and perception (Desirable)

* Experience in speech recognition (Desirable)

Skills & Competencies

* Good knowledge of written and spoken English.

Benefits

* Opportunity to work with a world-class inter-disciplinary speech research group.

To apply, please email a brief cover letter and CV, including the names and addresses of two

academic referees, to: kanejo@tcd.ie and to cegobl@tcd.ie

Back

Top

Nous sommes à la recherche d'une personne pour un stage de Master 2 recherche qui sera suivi d'un financement CIFRE de 3 ans avec AIRBUS. Merci de me contacter en m'envoyant un CV et une lettre de motivation le plus rapidement possible.

Contact
~~~~~~
Jérôme Farinas
Equipe SAMOVA
Institut de Recherche en Informatique de Toulouse
Tél : 05 61 55 74 34
Mèl : jfarinas@irit.fr

Mots clés
~~~~~~~
Parole spontanée, son, bruit, audio, reconnaissance, transcription, apprentissage.

Contexte de l’étude
~~~~~~~~~~~~~~~
Lors d’un vol, tous les paramètres sont enregistrés dans deux calculateurs distincts, le DFDR (Digital Flight Data Recorder) et le CVR (Cockpit Voice Recorder). Le DFDR enregistre les paramètres techniques du vol. Le CVR enregistre toutes les conversations entre l’équipage, le personnel de cabine, les centres au sol (navigation aérienne, compagnie…). Sont également enregistrés toutes les alarmes qui peuvent survenir à bord ainsi que tous les bruits qui peuvent être entendus dans le poste de pilotage.

Intérêt de l’étude
~~~~~~~~~~~~~
Au sein du département avionique, le contenu du CVR, suite à un vol d’essais et plus particulièrement ceux liés à une certification, est analysé afin de corréler tous les évènements tracés par les pilotes et ingénieurs navigants avec le contenu de l’enregistrement. Cette analyse a aussi pour but d’identifier et de caractériser tous les évènements sonores imprévus. L’analyse et la transcription sont faites par une simple écoute.
Afin d’améliorer la fiabilité, la pertinence, l’exhaustivité et la répétitivité des analyses effectuées, l’intérêt de l’étude est de proposer des algorithmes capables d’extraire du bruit ambiant du poste de pilotage, de la voix, des sons synthétiques et des bruits caractéristiques afin de les transcrire automatiquement.

Principaux objectifs de l’étude
~~~~~~~~~~~~~~~~~~~~~~~~
Ce stage adresse à la fois des problématiques liées à la reconnaissance de sons prédéfinis (existence d’une référence audio), la détection de bruits spécifiques et la transcription de la parole spontanée.
Ce sujet de stage est lié à une thèse qui sera financée avec un financement CIFRE avec AIRBUS. La thèse est divisée en 3 parties correspondant à l’étude de chacune des trois catégories de sons explorées ; les approches seront de fait différentes :
1. Pour la reconnaissance des sons (alarmes, code morse…), il est possible de définir des prototypes ou références. C’est pourquoi l’étude s’orientera vers l’emploi d’une part d’algorithmes de détection de fréquences caractéristiques en prenant en compte les contraintes d’un environnement hétérogène (bruit, recouvrement…) et d’autre part de méthodes de reconnaissance des formes classiques appliqués à l’audio tels que ceux employés en reconnaissance de motifs sonores [13].
2. La détection des bruits représentatifs (régime moteur, train d’atterrissage…) ou inattendus (usure anormale, parasite…), passera par une analyse des signatures
acoustiques caractéristiques pour en déduire une modélisation. Cette détection nécessitera une phase d’apprentissage.
Pour ces deux premières parties, le titulaire s’appuiera sur les résultats d’un stage de fin de cycle d’ingénieur pour lequel un prototype a déjà été développé. Il s’agira
d’enrichir la démarche scientifique, puis de compléter et de confirmer les choix techniques qui ont été proposés.
3. La reconnaissance de la parole qui est de loin la plus importante, s’orientera d’abord vers une étude de faisabilité. Les deux principales difficultés sont liées à la production même de la parole qui est quasi « spontanée » et à l’environnement dans lequel la prise de son est faite.
Dans un premier temps, le titulaire devra effectuer une analyse très précise de l’environnement, bien plus contraignant que les environnements « conventionnels »
dans lesquels sont développés la majorité des systèmes de reconnaissance de la parole (parole téléphonique, journaux d'information en anglais [1], français [2],
sessions du parlement européen [3]). Il existe des études sur l'influence de la dégradation due aux environnements bruités : dans le secteur de la construction [4],
dans le secteur aéronautique [5]. Les travaux dans ce dernier secteur ont principalement pour objectif depuis les années 1980 de réaliser une simple
commande vocale dans les cockpits. L'analyse de la parole à partir d'enregistrements CVR n’a jamais fait l'objet d'étude, seules des recherches sur les sons d'alerte ont été menées [6,7].
Dans un second temps et une fois l’analyse effectuée, le titulaire s’attachera à définir les spécifications du système de reconnaissance automatique, tout en respectant les contraintes suivantes :
- Contraintes liées à un environnement fortement bruité et altéré par des événements sonores liés à l’exploitation de l’avion (recouvrement).
- Contraintes liées à une population multiculturelle : langue (capacité à changer de langue pour un même locuteur), accent, vocabulaire.
- Contraintes liées aux conditions dans lesquelles le locuteur opère : augmentation du débit, stress, fatigue.
Les principales pistes de recherche envisagées se déclinent selon deux axes :
- La compensation au niveau de la paramétrisation : une première analyse des différents bruits des enregistrements CVR permettra de cibler les traitements pour lutter contre le bruit existant (soustraction cepstrale, normalisation de la variance, filtrage ARMA, filtres RASTA...). Un deuxième axe consiste à s’appuyer sur des résultats récents en compensation de bruit dans le domaine cepstral, domaine de paramétrisation le plus performant en reconnaissance de parole. Il s’agit de décomposer l’espace de représentation en une composante utile pour la reconnaissance et une composante dite de nuisance qui rend compte de la variabilité de la session d’enregistrement [14]. Dans le cas présent et le cadre particulier des cockpits, cette variabilité représenterait l’environnement bruité, mais elle pourrait aussi rendre compte des conditions de stress du locuteur.
- L’adaptation au niveau des modèles acoustiques et linguistiques : les modélisations de la parole devront être adaptées aux différentes formes de parole présentes dans les enregistrements. Il s’agira de regarder plus précisément les vocabulaires employés pour les enrichir éventuellement, et tenir compte de la spontanéité au travers des modèles de langage (le traitement de la spontanéité reste à ce jour un défi très important à relever). Les techniques d'adaptation des lois des modèles acoustiques par Maximum Likelihood Linear Regression (MLLR) [8] et Maximum A Posteriori (MAP) [9] et les techniques d’adaptation au niveau de la modélisation elle-même (analyse factorielle appliquée aux modèles de Markov cachés [10,11]) seront les points de départs. Les adaptations des lexiques et des modèles de langage devront être étendues à ce type de dialogues [12].
Durant le stage, les parties 1 et 2 devront donner lieu à un prototype utilisable. La partie 3 pourra être concrétisé par la mise en place d’un système de reconnaissance de la parole de base.

Back

Top

Ph.D. Researcher in Speech Synthesis, Trinity College, Dublin, Ireland

Post Specification

Post Title:

Ph.D. Researcher in Speech Synthesis

Post Status:

3 years

Department/Faculty:

Centre for Language and Communication Studies (CLCS)

Location:

Phonetics and Speech Laboratory

Salary:

€16,000 per annum (plus fees paid)

Closing Date:

31st January 2013

Post Summary

A Ph.D. Researcher is required to work in the area of speech synthesis at the Phonetics and

Speech Laboratory, School of Linguistic, Speech and Communication Sciences. The position

will involve carrying out research on the topic of Hidden Markov Model (HMM)-based speech

synthesis. Specifically, we are looking for a researcher to work on developing a source-filter

based acoustic modelling for HMM-based speech synthesis which is closely related to the

human speech production process and which can facilitate modification of voice source and

vocal tract filter components at synthesis time.

Background to the Post

Much of the research carried out to date in the Phonetics and Speech Laboratory has been

concerned with the role of the voice source in speech. This research involves the development

of accurate voice source processing both as a window on human speech production and for

exploitation in voice-sensitive technology, particularly synthesis. The laboratory team is interdisciplinary

and includes engineers, linguists, phoneticians and technologists.

This post will the main be funded by the on-going Abair project which has developed the first

speech synthesisers for Irish (

www.abair.ie), and the researcher will exploit the current Abair

synthesis platform. In this project the aim is to deliver multi-dialect synthesis with multiple

personages and voices that can be made appropriate to different contexts of use. The post will

also be linked to the FastNet project which aims at voice-sensitive speech technologies. A

specific goal of our laboratory team is to leverage our expertise on the voice by improving the

naturalness of parametric speech synthesis, as well as making more flexible synthesis platforms

which can allow modifications of voice characteristics (e.g., for creating different

personalities/characters, different forms of expression etc).

Standard duties of the Post

Initially the researcher will be required to attend some lectures as part of the Masters

programme on Speech and Language Processing. This and a supervised reading

programme will provide a background in the area of voice production, analysis and

synthesis.

Back

Top

15 RESEARCH POSITIONS (PhD, Post-Doc and Research Programmer) IN
MACHINE TRANSLATION, PARSING, INFORMATION RETRIEVAL, INFORMATION
EXTRACTION AND TEXT ANALYTICS AT CNGL, DCU. CLOSING DATE 15th FEBRUARY
2013
-------------

At the Centre for Next Generation Localisation (CNGL, www.cngl.ie) at
Dublin City University (DCU), Ireland, we are recruiting 11 PhDs, 3
Post-Doctoral Researchers and 1 Research Programmer (1 position).

-------------

CNGL is a €50M+ Academia-Industry partnership, funded jointly by
Science Foundation Ireland (SFI) and our industry partners, and is
entering its second cycle of funding. CNGL is looking to fill multiple
posts associated with its second phase which will focus on expansion
of our work into the challenging areas of social text sources and
multimedia content.

CNGL is an active collaboration between researchers at Dublin City
University (DCU), Trinity College Dublin (TCD), University College
Dublin (UCD), University of Limerick (UL), as well as 10 industrial
partners, including SMEs, Microsoft, Symantec, Intel, DNP, and
Welocalize.

CNGL comprises over 100 researchers across the various institutions
developing novel technologies addressing key challenges in the global
digital content and services supply chain. CNGL is involved in a large
number of European FP7 projects, as well as commercial projects in the
areas of language technologies, information retrieval and digital
content management. CNGL provides a world class collaborative research
infrastructure, including excellent computing facilities, and
administrative, management and fully integrated and dedicated on-site
commercialisation support.

The successful candidates will become part of the research team based
at DCU, joining two leading academic MT/NLP/IR and Translation
research groups (www.nclt.dcu.ie/, cttsdcu.wordpress.com/). The team’s
location at DCU, minutes from Dublin city centre, offers a highly
conducive environment for research, collaboration and innovation with
a wealth of amenities on campus.

DCU is ranked in the TOP 50 of young universities worldwide (under 50
years old) (QS Ranking) and in the TOP 100 under the Times Higher
Education (under 50 years) ranking scheme.

The research is supervised by Dr. Jennifer Foster, Dr. Sharon O'Brien,
Dr. Gareth Jones, Prof. Qun Liu and Prof. Josef van Genabith.

-----------------------------------
--* CNGL PHD STUDENSHIPS AT DCU *--
-----------------------------------

Parsing, Analytics and Information Extraction:

[PhD_CC_1] Tuning Text Analytics to User-Generated Content: Parse
quality estimation and targeted self-training.
[PhD_CC_2] Extracting Events and Opinions from User-Generated Content:
Deep parsing-based methods.

Information Retrieval:

[PhD_SD_1] Self-Managing Information Retrieval Technologies: Query,
search technique and parameter selection in information retrieval
applications
[PhD_SD_2] Indexing and Search for Multimodal (Spoken/Visual) Content:
Locating relevant content in multimodal sources

[PhD_SDCC_1] Application of Text Analytics in Information Retrieval:
Enhancing information retrieval using features from text analysis
[PhD_SDDI_1] Investigating Human-Computer Interaction Issues for
Search and Discovery with Multimodal (spoken/Visual) Content

Machine Translation:

[PhD_TL_1] Syntax- and Semantics-Enhanced Machine Learning Based MT
[PhD_TL_2] Domain Adaptation Based on Multi-Dimensional Quality
Estimation, Similarity Metrics, Clustering and Search
[PhD_TL_3] Human interaction with MT output: Usability, Acceptability,
Post-editing Research

[PhD_TLDI_1] MT and Multimodal Interaction
[PhD_TLSD_1] MT for Multimodal Cross Language Information Retrieval

Ideal candidates for the PhD studentships (except PhD_TL_3 - see
below) should have:

 - excellent computing and mathematical skills
 - strong machine learning and statistical skills
 - strong interest in basic research, applied research and showcasing
research in demonstrator systems
 - willingness to work as a team, but also be able to work on their initiative
 - strong ability in independent and creative thinking
 - strong problem-solving skills
 - excellent communication abilities (including writing and
presentation skills)
 - proficiency in English
 - a background in NLP, Computational Linguistics, Information
Retrieval, Information Extraction or Machine Translation as
appropriate for the relevant position is an advantage

Ideal candidates for the PhD studentship PhD_TL_3 (human interaction
with MT output: Usability, Acceptability, Post-editing Research)
should have:

 - excellent skills in empirical research in translation, technical
communication and/or HCI
 - strong interest in basic research, applied research and showcasing
research in demonstrator systems
 - willingness to work as a team, but also be able to work on own initiative
 - strong ability in independent and creative thinking
 - strong problem-solving skills
 - excellent communication abilities (including writing and
presentation skills)
 - high proficiency in English
 - a background in Translation, NLP, Computational Linguistics, or
Machine Translation is an advantage

PhD positions are fully funded for 3 years. Stipend: Fees + €16,000
p.a. living expenses (tax free)

-------------------------------
--* POST-DOCTORAL POSITIONS *--
-------------------------------

Parsing, Analytics and Information Extraction:

[PD_CC_1] Extracting Events and Opinions from User-Generated Content:
Parsing-based deep methods (up to 2 year contract)
[PD_CC_2] Extracting Events and Opinions from UGC: Shallow methods,
including unsupervised methods (up to 2.5 year contract)

Machine Translation:

[PD_TL_1] User/Human Centric MT (up to 2.5 year contract)

Ideal candidates for the post-doctoral positions should have:

 - a strong international publication record
 - a background in NLP, Computational Linguistics, Information
Retrieval, Information Extraction or Machine Translation as
appropriate for the position
 - excellent computing and mathematical skills
 - strong machine learning and statistical skills
 - strong interest in basic research, applied research and showcasing
research in demonstrator systems
 - willingness to work as a team, but also be able to work on their
own initiative
 - strong ability for independent and creative thinking
 - strong problem-solving skills
 - excellent communication abilities (including writing and
presentation skills)
 - proficiency in English
 - ability to identify and develop future research and funding initiatives
 - willingness to supervise and assist undergraduate and postgraduate students
 - ability to lead small teams of researchers, in co-operation with
the principal investigator


Indicative starting salary (subject to experience and qualifications):
 €37,750 - €42,394 (taxable)

-------------------------------------
--* RESEARCH PROGRAMMER POSITIONS *--
-------------------------------------
[RProg_CC_1] Research Programmer (up to 2.5 year contract)

Parsing, Analytics and Information Extraction:

Candidates for the researcher programmer position will have:
 - strong computer engineering and design skills
 - excellent knowledge of one or more of the following: Java, C++, PHP, Python
 - comfortable working in both UNIX and Windows environments
 - excellent algorithmic and analytical skills
 - candidates will hold an MSc/PhD in Computer Science/Software Engineering
 - excellent communication abilities
 - willingness to work as part of a team, but also be able to work on
your own initiative
 - experience in Natural Language Processing, Artificial
Intelligence, Information Retrieval, Localisation etc. would be highly
advantageous
 - experience with research-based software development and/or
cloud-based and grid computing technologies is also highly preferred


Indicative Salary (subject to experience and qualifications):
€37,750 - €42,394 (taxable)

-------------

**CLOSING DATE FOR APPLICATIONS ALL POSITIONS: 15th FEBRUARY 2013**
For more details please see: http://www.cngl.ie/vancancies.html

Application forms are available from:
http://www.dcu.ie/vacancies/APPLICATION_FORM_8pg.doc. Please also send
a CV with two contact points for references. When completing your
application, please indicate which positions you are applying for (in
order of preference) e.g. [PHD_TL_1], [PHD_TL_2]. Completed
application forms should be sent to Dr. Declan Groves
.


*For informal enquiries please contact the relevant PI below*:

- Dr. Jennifer Foster     [PhD_CC_1]
- Prof. Josef van Genabith
[PhD_CC_2],[PD_CC_1],[PD_CC_2],[RProg_CC_1]
- Dr. Gareth Jones     [PhD_SD_1],
[PhD_SD_2], [PhD_SDCC_1], [PhD_SDDI_1]
- Prof. Qun Liu     [PhD_TL_1], [PhD_TL_2],
[PhD_TLDI_1], [PhD_TLSD_1], [PD_TL_1]
- Dr. Sharon O'Brien    [PhD_TL_3]

Back

Top

This is a pre-announcement for a position in the Computer Science Department at the Cooperative State University in Karlsruhe, Germany, for a

Ph.D. Research Assistant or Post-Doctoral Researcher

in the field of Automatic Language Processing for Education

to be filled immediately with a salary according to TV-L E13 at 50% for 18 Months. The opening is in Karlsruhe, Germany, as part of a joint research project between Karlsruhe Institute of Technology (KIT), the Cooperative State University (DHBW) and the University of Education (PH) sponsored by DFG involving speech technology for educational system. (Working language: English or German)

Project cooperation partners are:

Cooperative State University (Duale Hochschule, Karlsruhe)

University of Education, Karlsruhe (Pädagogische Hochschule, Karlsruhe)

Karlsruhe Institute of Technology (KIT)

Description

Starting as soon as possible, we are seeking an experienced and motivated person to join our team of researchers from the above mentioned institutes. The ideal candidate will have knowledge in computational linguistics and algorithm design. Responsibilities include the use and improvement of research tools to update and optimize algorithms applied to diagnostics in children’s (German) writing using speech recognition and speech synthesis tools. For further details of this work, please refer to publications at SLaTE 2011, Interspeech 2011, and WOCCI 2012 by authors Berkling, Stüker, and Fay. Joint and collaborative research between the partners will be very close, offering exposure to each research lab. 

Candidates:

• Doctoral Research Candidates may apply and are welcome for joint research with their host institution.

• Experienced (post-doctoral) Research Candidates are already in possession of a doctoral degree or have at least 3 years but less than 5 years of research experience in engineering and/or hearing research.

Requirements

• Higher degree in speech science, linguistics, machine learning, or related field

• Experience developing ASR applications - training, tuning, and optimization

• Software development experience (for example: Perl, TCL, Ruby, Java, C)

• Excellent communication skills in English

• Willingness and ability to spend 18 months in Germany, working in a team with project partners

• Knowledge of German linguistics, phonics, graphemes, morphology or willingness to learn

• Strong interest in computational linguistics, morphology, phonics for German

Desirable:

• Interest in Education and language learning

• Interest in Human Computer Interaction and game mechanics

• Ability to create Graphic interfaces in multi-player applications

Application Procedure: Non-EU candidates need to check their ability to reside in Germany. Interested candidates, please send application (CV, certified copies of all relevant diplomas and transcripts, two letters of recommendation , proof of proficiency in English, letter of motivation (research interest, reason for applying to position) to: Berkling@dhbw-karlsruhe.de

The Cooperative State University is pursuing a gender equality policy. Women are therefore particularly encouraged to apply. If equally qualified, handicapped applicants will be preferred.

Back

Top

-- Research Associate in Robust Speech Recognition
===============================================

The University of Sheffield, Department of Computer Science invite applications for a position as Research Associate to work on a project to research and develop robust technology for recognition of speech in meetings.  The associated project, DocuMeet, is funded by the European Union and involves collaboration with partners from academia and industry across Europe. The Speech and Hearing Research Group is responsible for speech technology in the project, but also contributes to some natural language understanding tasks.

Speech transcription of meeting data is a well-established task, with international competitions (held by U.S. NIST), and supported by several large-scale research projects. While significant progress has been made, the performance of recognition, detection and analysis systems is still very far from usable in many realistic, natural scenarios. There are many significant challenges seeking a solution: the acoustic complexity of meetings goes well beyond standard settings.: Noise and reverberation are standard; speech signals show significant amounts of overlap between speakers; varying degrees of emotion are present; and speakers are moving. All of these pose significant challenges to speech research and practical applications.

In the DocuMeet project we specifically work on speech recognition robustness to noise and reverberation. We aim to work on new algorithms that allow to factor environment and context in novel ways (e.g. eigen-environments).  The recordings from multiple microphones can be used to remove unwanted acoustics, while knowledge about a specific environment type should be used to adjust acoustic models of the recognition systems. Further we will investigate how such algorithms can be integrated with personalisation (acoustic/language) and how metadata can be used to inform such processes. Extensive experimentation of existing and new corpora will be required to demonstrate the effectiveness of the new techniques.

Applicants are required to have a track record of work on speech technologies including speech recognition, and to have had exposure to modern machine learning techniques. Ideally, such a track record is demonstrated by publications in international journals and conferences. The successful candidate will be required to hold a PhD in the field; work on the project will require publication of results, travelling to conferences and extensive visits to itslanguage offices. At this point the project duration is for one year, but extensions are likely.
The project will be embedded in the Speech and Hearing (SpandH) research group at (http://www.dcs.shef.ac.uk/spandh) in the Department of Computer Science, and in particular the subgroup on machine intelligence for natural interfaces (MINI). SpandH is amongst the largest speech research groups in the UK, with extensive infrastructure and a vibrant research environment. The group is well known internationally for its research, which reaches across traditional divides to encompass and link computational hearing, speech perception and speech technology. The MINI subgroup is led by Prof. Hain, currently has 13 members, and is amongst other things well known for speech recognition and classification. It has had systems with best performance in international competitions that are available to the public at www.webasr.org. The subgroup is currently involved in many projects, including an EPSRC programme grant (with Univ. of Cambridge, Univ. of Edinburgh), research organisations (e.g. Idiap, NICT), and Industry (e.g. Cisco, Google). It has its own extensive computing infrastructure, access to large quantities of data, as well as dedicated recording facilities.

The Department of Computer Science, which is a member of the Faculty of Engineering, was established in 1982 and has since attained an international reputation for its research and teaching. Currently there are over 100 members of staff in Computer Science, including 35 Academics. The Department has an international reputation for the quality of its research, and was awarded grade 5 in the 2001 research assessment exercise, and in the 2008 exercise, 65% of our research was rated world leading or internationally excellent in terms of its originality, significance and rigor.

If you would like to know more about this position, please contact Prof. Thomas Hain - t.hain@dcs.shef.ac.uk.

In order to apply, the best option is to visit jobs.ac.uk and then press the 'Apply' button on the page:

  http://www.jobs.ac.uk/job/AFU678/research-associate/

The University of Shefield JOB ID is UOS005891.

Back

Top

Position: Developer for Large-Scale Audio Indexing Technologies: 1 W/M position at IRCAM

Starting:  March 2013

Duration: 18 months

Position description

He/she will be in charge of the development of a framework for scalable storage, management and access of distributed data (audio and meta-data). He/she will be also in charge of the development of scalable search algorithms.

Required profile:

·       High skill in database management systems

·       High skill in scalable indexing technologies (hash-table, m-trees …)

·       High skill in C++ development (including template-based meta-programming)

·       Good knowledge of Linux, Mac OSX and Windows development environment (gcc, Intel and MSVC, 
svn)

·       High productivity, methodical work, excellent programming style.

The developer will collaborate with the project team and participate in the project activities (evaluation of technologies, meetings, specifications, reports).  

Introduction to IRCAM

IRCAM is a leading non-profit organization associated to Centre Pompidou, dedicated to music production, R&D and education in sound and music technologies. It hosts composers, researchers and students from many countries cooperating in contemporary music production, scientific and applied research. The main topics addressed in its R&D department include acoustics, audio signal processing, computer music, interaction technologies, and musicology. Ircam is located in the center of Paris near the Centre Pompidou, at 1, Place Igor Stravinsky 75004 Paris.

Salary

According to background and experience

Applications

Please send an application letter together with your resume and any suitable information addressing the above issues preferably by email to: peeters_at_ircam_dot_fr with cc to vinet_a_t_ircam_dot_fr, roebel_at_ircam_dot_fr.

Back

Top

Proposal for an INRIA PhD fellowship (Cordi-S)

Title of the proposal: Nonlinear speech analysis for differential diagnosis between Parkinson's disease and Multiple-System Atrophy

Project Team INRIA: GeoStat (http://geostat.bordeaux.inria.fr/)

Author of the proposal research subject: Khalid Daoudi (khalid.daoudi@inria.fr)

Keywords: speech processing, nonlinear speech analysis, machine learning, voice pathology,

dysphonia, dysarthria, Multiple-System Atrophy, Parkinson's disease.

Scientific context:

Parkinson's disease (PD) is the most common neurodegenerative disorder after Alzheimer's disease.

Prevalence is 1.5% of the population over age 65 and affects about 143,000 French. Given the aging of the population, the prevalence is likely to increase over the next decade.

Multiple-System Atrophy (MSA) is a rare and sporadic neurodegenerative adult disorder, of

progressive evolution and of unknown etiology. The MSA has a prevalence of 2 to 5/100 000 and has no effective treatment. It usually starts in the 6th decade and there is a slight male predominance. It takes 3 years on average from the first signs of the disease for a patient to require a walking aid, 4-6 years to be in a wheelchair and about 8 years to be bedridden.

The PD and MSA require different treatment and support. However, the differential diagnosis between PD and MSA is a very difficult task because, at the early stage of the diseases, patients look alike as long as signs, such as dysautonomia, are not more clearly installed for MSA patients. There is currently no valid clinical nor biological marker for clear distinction between the two diseases at an early stage.

Goal:

Voice and speech disorders in Parkinson's disease is a clinical marker that coincides with a motor

disability and the onset of cognitive impairment. Terminology commonly used to describe these

disorders is dysarthria [1].

Like PD patients, depending on areas of the brain that are damaged, people with AMS may also have speech disorders: difficulties of articulation, staccato rhythm, squeaky or muted voice. Dysarthria in AMS is more severe and early in the sense that it requires more early rehabilitation compared to PD.

Since dysarthria is an early symptom of both diseases, the purpose of this thesis is to use dysarthria,

through digital processing of voice recordings of patients as a mean for objective discrimination

between PD and MSA. The ultimate goal is to develop a numerical dysarthria measure, based on the

analysis of the speech signal of the patients, which allows objective discrimination between PD and

MSA and would thus complement the tools currently available to neurologists in the differential

diagnosis of the two diseases.

Project:

Pathological voices, such as in PD and MSA, generally present high non-linearity and turbulence.

Nonlinear/turbulent phenomena are not naturally suited to linear signal processing. The latter is

however ruling over current speech technology. Thus, from the methodological point of view, the goal of this thesis is to investigate the framework of nonlinear and turbulent systems, which is better suited to analyzing the range of nonlinear and turbulent phenomena observed in pathological voices in general [2], and in PD and MSA voices in particular. We will adopt an approach based on novel nonlinear speech analysis algorithms recently developed in the GeoStat team [3]. The goal being to extract relevant speech features to design new dysarthria measures that enable accurate discrimination between

PD and MSA voices. This will also require investigation of machine learning theory in order to develop robust classifiers (to discriminate between PD and MSA voices) and to make correspondence (regression) between speech measures and standard clinical rates.

The PhD candidate will actively participate, in coordination with neurologists from the Parkinson's

Center of Haut-Lévêque Hospital, to set up the experimental protocol and data collection. The latter

will consist in recording patient's voices using DIANA or EVA2 workstation (http://www.sqlab.fr/).

References:

[1] Auzou, P.; Rolland, V.; Pinto, S., Ozsancak C. (eds.). Les dysarthries. Editions Solal. 2007.

[2] Baghai-Ravary L. ; Beet S.W. Automatic Speech Signal Analysis for Clinical Diagnosis and Assessment of Speech Disorders. Springer 2013.

[3] PhD thesis of Vahid Khanagha. GeoStat team, INRIA Bordeaux-Sud Ouest. January 2013.

http://geostat.bordeaux.inria.fr/images/vahid%20khanagha%204737.pdf

Advisor: K. Daoudi

Duration: 3 years (starting fall 2013)

Prerequisites: Good level in signal/speech processing is necessary, as well as Matlab and C/C++

programing. Knowledge in machine learning would be a strong advantage.

Back

Top

Ph.D Student in Speech and Music Communication  

Back

Top

Un poste de Maître de conférences en informatique est ouvert au concours à l'Université Paris-Sud, à destination de l'IUT d'Orsay.

Le traitement de la parole (de l'interaction vocale à l'indexation de documents multimédia) fait partie des thématiques de recherche prioritaires du poste ; ce sujet est développé au LIMSI dans le groupe TLP (voir aussi http://www.limsi.fr/tlp/postes13.html).

Pour consulter la fiche détaillée du poste :

http://www.u-psud.fr/_attachments/enseignants_chercheurs/Fiche%2520emploi%252027%2520MCF%25201905x.pdf?download=true

et les modalités pour le dossier de candidature (à déposer avant le 28 mars 2013) :

http://www.u-psud.fr/fr/recrutement/enseignants/enseignants_chercheurs.html

Back

Top

Dans le cadre du projet ANR ContNomina(2013-2016), nous proposons

Back

Top

My department is now announcing four (not much, but still) positions for doctoral students. The positions mean four years fully financed (full time salary on the order of € 2500 per month), office space, computer and other technical resources necessary to do the job. You will find more information here:

http://www.flov.gu.se/english/education/doctoral-studies-third-cycle/admission/?languageId=100001&contentId=-1&disableRedirect=true&returnUrl=http%3A%2F%2Fwww.flov.gu.se%2Futbildning%2Fforskarniva%2Fansokan%2F

I would greatly appreciate if you could make this opportunity known among your students.

Anders Eriksson, MSc, PhD.
Professor of Phonetics
Department of Philosophy, Linguistics and Theory of Science
University of Gothenburg
Box 200, SE-405 30 Gothenburg, SWEDEN
Web: http://www.ling.gu.se/~anders <http://www.ling.gu.se/%7Eanders>

Back

Top

Job offer: Post-doc
    A post-doctoral position in linguistics is available at François     Rabelais University, Tours (France), to work with the linguists in     the Inserm Research Unit 930 Imagerie & cerveau on a research     project funded by the French National Research Agency, BiLaD     (Développement du langage bilingue : Enfants à développement typique     et enfants avec troubles du langage).
    Starting date: September 1st, 2013 for a duration of 12 months     (renewable for a total of 24 months).
    To apply, send a CV + names and contact information for two     recommenders to Laurie Tuller (tuller@univ-tours.fr).
    Responsibilities: Organization, processing, and analysis of     linguistic and psycholinguistic data from experimental and     standardized tasks (verbal and non-verbal).
    Required training and skills: PhD in linguistics or related field     (psychology, cognitive science) with specialty in language     acquisition. Solid experience in transcribing linguistic data, and     in use of data entry and processing software (Excel, SPSS, PRAAT).
    Other skills: Experience with children with SLI, familiarity with     the field of speech-language pathology/therapy and bilingualism.     Ability to work in French and in English required; mastery of     Turkish or Portuguese would be an asset.
    Deadline : May 15, 2013.

Back

Top

Poste de Maître de Conférences en Informatique
Profil 'Traduction automatique statistique'
Université du Maine, Le Mans
Référence GALAXIE : 4055


Profil recherche : 

Le candidat doit avoir des connaissances approfondies dans au moins un des domaines suivantes : la traduction automatique statistique, le traitement automatique de la langue, l'apprentissage automatique. Une expérience dans la construction de grands systèmes de traduction avec l'outil Moses est particulièrement bienvenue.


Les candidats sont invités à se manifester auprès de :
Holger Schwenk : holger.schwenk@lium.univ-lemans.fr

Lieu(x) d’exercice : LIUM, Université du Maine, site du Mans
Nom directeur labo : Yannick Estève
Tel directeur labo : 02 43 83 3874
Email directeur labo : yannick.esteve@lium.univ-lemans.fr 
URL labo : http://www-lium.univ-lemans.fr


Descriptif du laboratoire :
Le Laboratoire d'Informatique de l'Université du Maine (LIUM - EA 4023), créé il y a environ vingt-cinq ans, regroupe la plupart des enseignants-chercheurs en informatique de l’Université du Maine. Il comprend actuellement environ 50 personnes : 20 enseignants-chercheurs, 8 post-doctorants, 13 doctorants et 4 BIATOSS. Le LIUM est composé de deux équipes : une équipe de neuf enseignants-chercheurs permanents spécialisés en Environnements informatiques pour l'apprentissage humain (Ingénierie des EIAH), dirigée par Christophe Choquet ; une équipe de onze enseignants-chercheurs spécialisés en reconnaissance de la parole et traduction automatique (Language and Speech Technology LST), dirigée par Paul Deléglise.


Profil enseignement : 

Le candidat devra être capable d'enseigner en français les matières fondamentales de l'informatique (algorithmique, programmation orientée objet, réseaux, bases de données, génie logiciel, etc).


Nom du responsable du Département (enseignement) : 
Christophe Després Christophe.Despres@univ-lemans.fr

Back

Top

Nuance Healthcare, a division of Nuance Communications, is the market leader in providing clinical understanding solutions that accurately capture and transform the patient story into meaningful, actionable information. Thousands of hospitals, providers and payers worldwide trust Nuance speech-enabled clinical documentation and analytics solutions to facilitate smarter, more efficient decisions across the healthcare enterprise. These solutions are proven to increase clinician satisfaction and HIT adoption, supporting organizations to achieve Meaningful Use of EHR systems and transform to the accountable care model. Nuance Healthcare has been recognized as “Best-in-KLAS” 2004-2012 for Speech Recognition.

Research Scientist (m/f) – Computational Linguist NLP

Preferred Vienna / Austria

As a Research Scientist - Computational Linguist - you will be part of the Healthcare Automatic Speech Recognition Research team. You will work on research and development of algorithms, resources and methods to support data collection and improve accuracy of Nuance healthcare products.    

Your Task.

• Development of methods, algorithms, resources and tools in the area of natural language processing (NLP) and computational linguistics
• Experimental and theoretical analysis of computational linguistics and NLP related problems
• Following academic development in speech recognition area, attending conferences and writing scientific papers (if applicable)

Your Profil.

• PhD or Master degree in (computational) linguistics, engineering sciences, mathematics or physics    
• Knowledge in machine learning methods (statistical, rule-based)
• Analytical and problem-solving skills
• Computer programming (C++ and/or scripting languages)
• Team player but also able to work on his own initiative
• Willing to learn
• English language    

Preferred

• Experience (industrial or academic) in NLP or computational linguistics
• Knowledge in automatic speech recognition
• Linguistic knowledge    

Our offer.

We offer a competitive compensation package and a casual yet technically challenging work environment. Join our dynamic, entrepreneurial team and become part of our fast growing track of continuing success.  Nuance is an Equal Opportunity Employer.

• Innovative products  
• full time &   permanent employment
• International   projects and international teams
• Development within   Nuance organization

Does Nuance speak to you?

Please apply via our Recruiting tool on our homepage https://jobs-nuance.icims.com/jobs/9447/job or via EMEAjobs@nuance.com  reference number 9447-Research Scientist - Comp Linguist NLP. Please provide CV, supporting documents and letter of motivation including preferred country, start date and salary expectations.

For more information visit us www.nuance.com.

°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°

Research Scientist (m/f) – Language Modeling

Preferred Vienna / Austria

As a Research Scientist you will be part of the Healthcare Automatic Speech Recognition Research team. You will work on research and development of ASR algorithms, resources and methods to improve accuracy and performance of Nuance healthcare products.

Your Task.

• Work in language modeling and grammar design area
• Development of training and adaptation tools and algorithms
• Performing speech recognition (computer) experiments and studies
• Experimental and theoretical analysis of speech recognition problems
• Following academic development in speech recognition area, attending conferences and writing scientific papers (if applicable)

Your Profil.

• PhD or Master degree in engineering sciences, mathematics or physics
• Knowledge in automatic speech recognition methods and algorithms
• Analytical and problem-solving skills
• Statistical data analysis
• Scripting languages in Unix/Linux environment (e.g. Perl, Python, Bash)
• Team player but also able to work on his own initiative
• Willing to learn
• English language

Preferred

• Experience (industrial or academic) in language modeling area
• Statistical pattern recognition
• Linguistic knowledge

Our offer.

We offer a competitive compensation package and a casual yet technically challenging work environment. Join our dynamic, entrepreneurial team and become part of our fast growing track of continuing success.  Nuance is an Equal Opportunity Employer.

• Innovative products  
• full time &   permanent employment
• International   projects and international teams
• Development within   Nuance organization

Does Nuance speak to you?

Please apply via our Recruiting tool on our homepage https://jobs-nuance.icims.com/jobs/9446/job or via EMEAjobs@nuance.com  reference number 9446-Research Scientist - LM. Please provide CV, supporting documents and letter of motivation including preferred country, start date and salary expectations.

For more information visit us www.nuance.com.

°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°

Speech Scientist (m/f) - Linguist

Preferred Vienna / Austria

As a Speech Scientist - (Computer) Linguist - you will be part of the Healthcare Automatic Speech Recognition Research team. You will work on research and development of resources, algorithms and methods to improve accuracy and performance of Nuance healthcare products.

Your Task.

• Development of rules (grammars), lexicons and statistic models for speech recognition
• Performing speech recognition (computer) experiments and studies
• Following academic development in speech recognition area, attending conferences and writing scientific papers (if applicable)

Your Profil.

• PhD or Master degree in linguistic, machine learning, computer linguistic or in a related field
• (Computer) linguistic and/or machine learning knowledge
• Scripting languages (e.g. Perl and/or Python)
• Analytical and problem-solving skills
• Team player but also able to work on his own initiative
• Communication / coordination skills
• Willing to learn
• English language

Preferred

• Experience in the area of automatic speech recognition
• Knowledge of context independent grammars
• Working in Unix/Linux and Windows environment

Our offer.

We offer a competitive compensation package and a casual yet technically challenging work environment. Join our dynamic, entrepreneurial team and become part of our fast growing track of continuing success.  Nuance is an Equal Opportunity Employer.

• Innovative products  
• full time &   permanent employment
• International   projects and international teams
• Development within   Nuance organization

Does Nuance speak to you?

Please apply via our Recruiting tool on our homepage https://jobs-nuance.icims.com/jobs/9445/job or via EMEAjobs@nuance.com  reference number 9445-Speech Scientist Linguist. Please provide CV, supporting documents and letter of motivation including preferred country, start date and salary expectations.

For more information visit us www.nuance.com.

°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°

Research Scientist (m/f) – Acoustic Modeling

Preferred Vienna / Austria

As a Research Scientist you will be part of the Healthcare Automatic Speech Recognition Research team. You will work on research and development of ASR algorithms, resources and methods to improve accuracy and performance of Nuance healthcare products.

Your Task.

• Work in acoustic modeling area
• Development of training and adaptation tools and algorithms
• Performing speech recognition (computer) experiments and studies
• Experimental and theoretical analysis of speech recognition problems
• Following academic development in speech recognition area, attending conferences and writing scientific papers (if applicable)

Your Profil.

• PhD or Master Degree in engineering sciences, mathematics or physics
• Knowledge in automatic speech recognition methods and algorithms
• Analytical and problem-solving skills
• Statistical data analysis
• Scripting languages in Unix/Linux environment (e.g. Perl, Python, Bash)
• Team player but also able to work on his own initiative
• Willing to learn
• English language

Preferred

• Experience (industrial or academic) in acoustic modeling area
• Statistical pattern recognition
• Signal processing

Our offer.

We offer a competitive compensation package and a casual yet technically challenging work environment. Join our dynamic, entrepreneurial team and become part of our fast growing track of continuing success.  Nuance is an Equal Opportunity Employer.

• Innovative products
• full time & permanent employment
• International projects and international teams
• Development within Nuance organization

Does Nuance speak to you?

Please apply via our Recruiting tool on our homepage https://jobs-nuance.icims.com/jobs/9444/job or via EMEAjobs@nuance.com  reference number 9444-Research Scientist - AM. Please provide CV, supporting documents and letter of motivation including preferred country, start date and salary expectations.

For more information visit us www.nuance.com.

 °°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°

Research Scientist (m/f) - Acoustic Modeling

Preferred Vienna / Austria

As a Research Scientist you will be part of the Healthcare Automatic Speech Recognition Research team. You will work on research and development in area of acoustic modeling to improve accuracy and performance of Nuance healthcare products.

Your Task.

• Work in acoustic modeling area (e.g. creation of acoustic models for various languages)
• Performing speech recognition (computer) experiments and studies
• Experimental and theoretical analysis of speech recognition problems
• Following academic development in speech recognition area, attending conferences and writing scientific papers (if applicable)

Your Profil.

• PhD or Master degree in engineering sciences, mathematics or physics
• Knowledge in automatic speech recognition methods and algorithms
• Analytical and problem-solving skills
• Statistical data analysis
• Scripting languages in Unix/Linux environment (e.g. Perl, Python, Bash)
• Team player but also able to work on his own initiative
• Willing to learn
• English language

Preferred

• Experience (industrial or academic) in acoustic modeling area
• Statistical pattern recognition
• Signal processing

Our offer.

We offer a competitive compensation package and a casual yet technically challenging work environment. Join our dynamic, entrepreneurial team and become part of our fast growing track of continuing success.  Nuance is an Equal Opportunity Employer.

• Innovative products  
• full time &   permanent employment
• International   projects and international teams
• Development within   Nuance organization

Does Nuance speak to you?

Please apply via our Recruiting tool on our homepage https://jobs-nuance.icims.com/jobs/9443/job or via EMEAjobs@nuance.com  reference number 9443-Research Scientist AM. Please provide CV, supporting documents and letter of motivation including preferred country, start date and salary expectations.

For more information visit us www.nuance.com.

 °°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°°

Research Scientist (m/f) – Automatic Speech Recognition

Preferred Vienna / Austria

As a Research Scientist you will be part of the Healthcare Automatic Speech Recognition Research team. You will work on research and development of ASR algorithms, resources and methods to improve accuracy and performance of Nuance healthcare products.

Your Task.

• Application specific speech recognition research
• Tuning and optimization of speech recognition algorithms
• Software implementation
• Analysis of application specific speech recognition issues
• General expertize in speech recognition area

Your Profil.

• PhD or Master degree   in engineering sciences, mathematics or physics
• Analytical and   problem-solving skills
• Statistical data   analysis
• Linux OS including   bash
• Team player but also   able to work on his own initiative
• Willing to   learn
• English   language
• Either

   
• strong background     in the area of automatic speech recognition    
• some experience in     software development

or

   
• strong background     in software development    
• some experience in     the area of automatic speech recognition

Preferred Skills:

• Experience with an   object-oriented programming language
• Experience with   scripting languages like Perl and Python
• Statistical pattern   recognition

Our offer.

We offer a competitive compensation package and a casual yet technically challenging work environment. Join our dynamic, entrepreneurial team and become part of our fast growing track of continuing success.  Nuance is an Equal Opportunity Employer.

• Innovative products  
• full time &   permanent employment
• International   projects and international teams
• Development within   Nuance organization

Does Nuance speak to you?

Please apply via our Recruiting tool on our homepage https://jobs-nuance.icims.com/jobs/8047/job or via EMEAjobs@nuance.com  reference number 8047-Research Scientist - ASR. Please provide CV, supporting documents and letter of motivation including preferred country, start date and salary expectations.

For more information visit us www.nuance.com.

ILONA ALEXANDRA HOLTZ

Recruiter - Employment Specialist DACH

Human Resources

Nuance Communications Deutschland GmbH

Site Ulm

Soeflingerstr. 100

D-89077 Ulm, Germany

Fon       +49 731 - 379 50 1166

Fax       +49 731 - 379 50 1106 (Zentrale)

Mobil     +49 170 56 15 235

WWW.NUANCE.COM The experience speaks for itself ™

Geschäftsführung/Director: Caroline Curtis, Todd Michael DuChene, Thomas L. Beaudoin

Sitz der Gesellschaft/Registered Office: Aachen

Registergericht/Court of Registration: Aachen

Reg. Nr.: HRB 16313

USt-ID/VAT: DE 264500438

This electronic transmission and any files transmitted with it are confidential. It is transmitted for the sole use of the person(s) to whom it is addressed. Any further distribution or copying is prohibited. If you receive this message in error, please inform the sender immediately, do not use it or disclose its contents and delete it from your system. Please note that Nuance cannot guarantee that the transmission will be secure or error-free.

Experience Nuance in the web:

DragonDrive!

http://www.youtube.com/watch?v=wYIwVP2JqL4

DragonNaturallySpeaking - Say It Your Way!

http://www.youtube.com/watch?v=RkiYr8aw5pE

or meet me on XING or LinkedIn

Back

Top

Speech Data Evaluator for French

Job title:

Speech Data Evaluator for French (multiple positions)

In Dublin.

Job description:

As a Speech Data Evaluator and a native speaker of French, you will be part of a team based in Dublin, processing large amounts of linguistic data and carrying out a number of tasks to improve the quality of Google’s speech synthesis and speech recognition in your own language.

This includes:

• annotating and classifying linguistic data

• labeling text for disambiguation, expansion, and text normalization

• providing phonetic transcription of lexicon entries according to given standards and using in-house tools

Job requirements:

• native speaker of French (with good command of the standard dialect) and fluent in English

• computer-literate (should feel comfortable using in-house tools)

• attention to detail

• good knowledge of orthography and grammar in French

• passion for language and a keen interest in technology

• good organizational skills

• a degree in a language-related field such as linguistics, language teaching, translation, editing, writing, proofreading, or similar

Project duration: 6-9 months (with potential for extension)

For immediate consideration, please email your CV and cover letter in English (PDF format preferred) with 'Speech Data Evaluator French' in the subject line.

Back

Top

Offre de post-doc au LIUM
Traitement Automatique du Langage
–
Correction de sorties d’OCR, traduction automatique statistique, modélisation du langage.
Offre de postdoc au sein du laboratoire d’Informatique de l’Université du Maine (LIUM)  dans le domaine de la correction orthographique par méthodes de traduction automatique statistique.
Résumé de l’offre –  Thématiques : traitement du langage naturel, application à la correction d’OCR et traduction automatique statistique. –  Lieu : LIUM (Le Mans), équipe LST (http ://www-lium.univ-lemans.fr/). –  Période : disponible dès maintenant pour une durée d’un an renouvelable.
Contexte Ce postdoc s’inscrit dans le projet PACTE ('investissement d’avenir'), porté par l’entreprise Diadeis,  et dont sont également partenaires l’équipe Alpage (INRIA et Paris 7), et les entreprises A2ia et Isako.  PACTE a pour objectif l’amélioration de la qualité orthographique des textes issus de différentes méthodes de capture textuelle.  L’accent est mis sur les sorties d’OCR (reconnaissance optique de caractères) sur des textes imprimés scannés,  mais concerne également des données obtenues par reconnaissance d’écriture manuscrite, par saisie manuelle, et par rédaction directe.  Les techniques qui seront utilisées sont à la fois statistiques et hybrides, faisant usage d’outils et de ressources de linguistique computationnelle.
Objectifs Vérification et correction des sorties d’OCR par des méthodes de modélisation statistique du langage.  Les systèmes OCR utilisés n’exploitent pas ou peu de connaissance sur la langue.  L’objectif est d’exploiter la modélisation de la langue afin combler ce manque.
Utilisation de la traduction automatique statistique pour la correction d’erreurs des sorties d’OCR.  La correction des sorties d’OCR peut être vue comme une tâche de traduction d’un texte erroné vers un texte correct.  Dans le cadre de l’OCR, le paradigme de traduction doit être adapté afin de prendre en compte les spécificités de la tâche.
Le cadre applicatif de ce travail est assez exceptionnel, avec l’exploitation d’une grande quantité de données  issue notamment du Bureau Européen des Brevets (EPO - European Patent Office) et du Journal Officiel de l’Union Européenne.
Profil recherché – Compétences en informatique : environnement Linux, C++, scripting, etc. ; – Connaissances en apprentissage automatique, linguistique computationelle. – Une expérience en traduction automatique statistique est un plus. Le postdoc se déroulera au sein de l’équipe LST du LIUM.  Le LIUM est connu au niveau international pour ses recherches dans le domaine de la traduction automatique statistique,  et possède de nombreuses collaborations avec des universités et entreprises en Europe et aux États-Unis.
Contacts Envoyer une lettre de motivation et un CV montrant vos compétences pour ce poste aux adresses suivantes : Loïc Barrault : loic.barrault@lium.univ-lemans.fr  Holger Schwenk : holger.schwenk@lium.univ-lemans.fr 

Back

Top

Postdoctoral position at LIMSI, France
Expressive Vocal Signal Analysis and Modelling

Content
During the post-production of a movie, as well as for video games, digital double rendering techniques are used to modify the actor’s performance without the need of playing the scene again. This project is aim toward characterising and reproducing the actor’s vocal personality. By creating a digital double of an actor, it will be possible to create new auditory scenes, as well as dubbing a movie – preserving the actor’s voice, habits and vocal personality.
We thus aim at characterising the expressive space of a given speaker, mainly in the prosodic domain, but also for its vocal quality and articulation peculiarities.
This postdoctoral position thus consist in implementing a set of voice analysis methods in order to observe the variations of vocal source and articulation in a given set of expressive performances. Databases presenting variations of vocal effort and expressive speech are already available and labelled. They will be analysed in order to model the expressive changes in tem of variation of the vocal signal. These models will be confronted to high-level descriptions of the expressive content of the databases.

Required competences
This research requires sound knowledge of signal processing applied to speech analysis. Skills in phonetic and/or linguistic are valuable.
Teamwork skills, and deadline oriented environment is mandatory.
Candidate with signal processing, computational linguistic or acoustic phonetic background will be taken into consideration.

Research team
This position is part of the French FUI ADN T-R project. It takes place at LIMSI-CNRS (www.limsi.fr) in the Audio & Acoustic group, in collaboration with partners from the industry.
LIMSI-CNRS is a research lab located in the University Paris-Sud campus, at Orsay, south of Paris. LIMSI is internationally renown for its works on speech processing. The Audio & Acoustic group focuses more specifically on speech analysis and synthesis, real-time audio processing and expressive sound.
This project will start as soon as possible (from may 2013), and is funded by the FUI agency for a maximum of 16 months. Gross monthly salary is approximately 2500€, on the basis of CNRS post-doctoral contract.

Supervision & contact:
This postdoctoral position will be supervised by Christophe d’Alessandro and Albert Rilliard.
Applications should be send to:    

Back

Top

1.

Marie Curie Research Fellow in Speech Synthesis and Speech Perception

'Using statistical parametric speech synthesis to investigate speech perception'

The Centre for Speech Technology Research (CSTR) 
University of Edinburgh

This is a rare opportunity to hold a prestigious individual fellowship in a world-leading research group at a top-ranked University, mentored by leading researchers in the field of speech technology. Marie Curie Experienced Research Fellowships are aimed at the most talented newly-qualified postdoctoral researchers, who have the potential to become leaders in their fields. This competitively salaried fellowship offers an outstanding young scientist the opportunity to kick-start his or her independent research career in speech technology, speech science or laboratory phonetics. 

This fellowship is part of the INSPIRE Network (http://www.inspire-itn.eu) and the project that the CSTR Fellow will spearhead involves developing statistical parametric speech synthesis into a toolbox that can be used to investigate issues in speech perception and understanding. There are excellent opportunities for collaborative working and joint publication with other members of the network, and generous funding for travel to visit partner sites, and to attend conferences and workshops.

The successful candidate should have a PhD (or be near completion) in computer science, engineering, linguistics, mathematics, or a related discipline. He or she should have strong programming skills and experience with statistical parametric speech synthesis, as well as an appropriate level of ability and experience in machine learning. The fellowship is fixed term for 12 month (to start as soon as possible). CSTR is a successful and well-funded group, and so there are excellent prospects for further employment after the completion of the fellowship.

The Marie Curie programme places no restrictions on nationality: applicants can be of any nationality and currently resident in any country worldwide, provided they meet the eligibility requirements set out in the full job description (available online - URL below). 

Salary:  GBP 42,054 to GBP 46,731 plus mobility allowance 

Informal enquiries about this position should be made to Prof Simon King (Simon.King@ed.ac.uk) or Dr Cassie Mayo (catherin@inf.ed.ac.uk). 


Apply online:

https://www.vacancies.ed.ac.uk/pls/corehrrecruit/erq_jobspec_version_4.jobspec?p_id=013062

Closing date: 10 Jun 2013



2.


An Open Position for Postdoctoral Research Associate in Speech Synthesis 

The Centre for Speech Technology Research (CSTR) 
University of Edinburgh

This post holder will contribute to our ongoing research in statistical parametric ('HMM-based') speech synthesis, working closely with Principal Investigators Dr. Junichi Yamagishi and Prof. Simon King, in addition to other CSTR researchers. The focus of this position will be to conduct research into methods for generating highly intelligible synthetic speech, for a variety of applications, in the context of three ongoing and intersecting projects in CSTR: 

The 'SALB' project concerns the generation of extremely fast, but highly intelligible, synthetic speech for blind children. This is a joint project with the Telecommunications Research Centre Vienna (FTW) in Austria, and is funded by the Austrian Federal Ministry of Science and Research. 

The 'Voice Bank' project concerns the building of synthetic speech using a very large set of recordings of amateur speakers (‘voice donors’) in order to produce personalised voices for people whose speech is disordered, due to Motor Neurone Disease. This is a joint project with the Euan MacDonald Centre for MND research, and is funded by the Medical Research Council. The main tasks will be to conduct research into automatic intelligibility assessment of disordered speech and to devise automatic methods for data selection from the large voice bank. 

The 'Simple4All' project is a large multi-site EU FP7 project led by CSTR which is developing methods for unsupervised and semi-supervised learning for speech synthesis, in order to create complete text-to-speech systems for any language or domain without relying on expensive linguistic resources, such as labelled data. The main tasks here will be to further the overall goals of the project, including contributing original research ideas. There is considerable flexibility in the research directions available within the Simple4All project and the potential for the post holder to form a network of international collaborators. 

The successful candidate should have a PhD (or be near completion) in computer science, engineering, linguistics, mathematics, or a related discipline. He or she should have strong programming skills and experience with statistical parametric speech synthesis. 

Whilst the advertised position is for 24 months (due to the particular projects that the post-holder will contribute to), CSTR is a stable, well-funded and successful group with a tradition of maintaining long-term support for ongoing lines of research and of building the careers of its research fellows. We expect to obtain further grant-funded research projects in the future. 

Informal enquiries about this position to either Dr. Junichi Yamagishi (jyamagis@inf.ed.ac.uk) or Prof. Simon King (Simon.King@ed.ac.uk). 

Apply Online: 
https://www.vacancies.ed.ac.uk/pls/corehrrecruit/erq_jobspec_version_4.jobspec?p_id=013063

Closing date: 10 Jun 2013

Back

Top

A job opening is to be filled as soon as possible as part of the 'Deutschen Forschungsgemeinschaft (DFG)' sponsored project at the Department of Computer Science for the duration of up to 18 months at 50% of employment at the Cooperative State University in Karlsruhe, Germany, for a

Ph.D. Research Assistant

or Post-Doctoral Researcher

in the field of

Automatic Language Processing for Education

arch A

This job opening is in the field of automatic speech recognition as part of a joint research project between Karlsruhe Institute of Technology (KIT), the Cooperative State University (DHBW) and the University of Education (PH) sponsored by DFG involving speech technology for educational system. (Working language: English or German)

Description

Starting as soon as possible, we are seeking an experienced and motivated person to join our team of researchers from the above mentioned institutes. The ideal candidate will have knowledge in computational linguistics and algorithm design. Responsibilities include the use and improvement of research tools to update and optimize algorithms applied to diagnostics in children’s (German) writing using speech recognition and speech synthesis tools. For further details of this work, please refer to publications at SLaTE 2011, Interspeech 2011, and WOCCI 2012 by authors Berkling, Stüker, and Fay.

Joint and collaborative research between the partners will be very close, offering exposure to each research lab.

Candidates

:

-

Doctoral Research Candidates may apply and are welcome for joint research with their host institution.

-

Experienced (post-doctoral) Research Candidates are already in possession of a doctoral degree or have at least 3 years but less than 5 years of research experience in engineering and/or hearing research.

Requirements

-

Higher degree in speech science, linguistics, machine learning, or related field

-

Experience developing ASR applications - training, tuning, and optimization

-

Software development experience (for example: Perl, TCL, Ruby, Java, C)

-

Excellent communication skills in English

-

Willingness and ability to spend 18 months in Germany, working in a team with project partners

-

Knowledge of German linguistics, phonics, graphemes, morphology or willingness to learn

-

Strong interest in computational linguistics, morphology, phonics for German

Desirable

:

-

Interest in Education and language learning

-

Interest in Human Computer Interaction and game mechanics

-

Ability to create Graphic interfaces in multi-player applications

-

Working with Ruby on Rails

The job will allow for interesting work within a modern and well equipped environment in the heart of Karlsruhe. The salary level, depending on your circumstances, will be in line with the 13 TV-L tarrifs. KIT and the Cooperative State University are pursuing a gender equality policy. Women are therefore particularly encouraged to apply. If equally qualified, handicapped applicants will be preferred (please submit your paperwork accordingly). Non-EU candidates need to check their ability to reside in Germany.

Interested candidates, please send application (CV, certified copies of all relevant diplomas and transcripts, two letters of recommendation, proof of proficiency in English, letter of motivation (research interest, reason for applying to position) with notification of the job-number to be received on or before

April 26, 2013.

Send electronic application to:

berkling@dhbw-karlsruhe.de.

Questions about details of the job can be directed to berkling@dhbw-karlsruhe.de.

Back

Top

Attention : le dossier de candidature complet devra être soumis sur le site de l’EDITE au plus tard le 22 mai

Sujet de thèse : Traitement du contenu verbal et analyse des sentiments dans les systèmes d’interactions humain-agent

Proposé par : Chloé Clavel
Directeur de thèse: Catherine Pelachaud
Encadrant : Chloé Clavel
Unité de recherche: UMR 5141 Laboratoire Traitement et Communication de l'Information
Domaine: Département Traitement du Signal et des Images
Secteur: Traitement Automatique du Langage Naturel, Dialogue Homme-Machine
Thématique P: Signal Image SHS
Financement : bourse EDITE (voir modalités http://edite-de-paris.fr/spip/spip.php?article172)

Personnes à contacter :
chloe.clavel@telecom-paristech.fr
catherine.pelachaud@telecom-paristech.fr

**Projet
Le domaine du sentiment analysis et de l’opinion mining  est un domaine en plein
 essor avec l’arrivée en masse de données textuelles sur le web comportant des
 expressions d’opinions par les citoyens (critiques de films, débats sur les
 commentaires de forums, tweets) (El-Bèze et al. 2010)). Les recherches en
 traitement automatique des langues se mobilisent sur le développement de 
méthodes de détection d’opinion dans les textes en s’appuyant sur ces nouvelles
 ressources. La diversité des données et des applications industrielles faisant
 appel à ces méthodes multiplient les défis scientifiques à relever avec,
 notamment, la prise en compte des différents contextes d’énonciation (e.g.,
 contexte social et politique, personnalité du locuteur) et la définition du 
phénomène d’opinion à analyser en fonction du contexte applicatif. Ces méthodes
 d’analyse des sentiments dans les textes s’étendent également depuis peu à
 l’oral en passant par l’analyse des transcriptions automatiques issues de 
systèmes de reconnaissance automatique de la parole pour des problématiques 
d’indexation d’émissions radiophoniques ou de centres d’appels (Clavel et al.,
 2013), et peuvent être ainsi corrélées aux méthodes d’analyse
 acoustique/prosodique des émotions (Clavel et al., 2010).

Autre domaine scientifique en plein essor, celui des agents conversationnels
 animés (ACA) fait intervenir  des personnages virtuels intéragissant avec
 l’humain. Les ACA peuvent prendre un rôle d’assistant comme les agents
 conversationnels présents sur les sites de vente (Suignard, 2010), de tuteur 
dans le cadre des Serious Games (Chollet et al. 2012) ou encore de partenaire
 dans le cadre des jeux vidéos. Le défi scientifique majeur pour ce domaine est
 l’intégration, au sein de l’ACA, de la composante affective de l’interaction. 
Il s’agit d’une part de prendre en compte les comportements affectifs et des 
attitudes sociales de l’humain et d’autre part de les générer de façon
 pertinente.

Nous proposons pour cette thèse de travailler sur la détection des opinions et
 des sentiments dans un contexte d’interaction multimodale de l’humain avec un
 agent conversationnel animé, sujet jusqu'à maintenant peu étudié par la
 “communauté agent”. En effet, d’un côté, les ACA réagissent à des contenus
 émotionnels essentiellement non verbaux (Schröder et al., 2011) et de l’autre
 côté, les ACA “assistant” réagissent à partir des contenus verbaux informatif
 (Suignard, 2010) sans prendre en compte les opinions ou les sentiments exprimés
 par l’utilisateur. Des premières études ont été réalisées sur la reconnaissance
 de l’affect dans le langage dans un contexte d’interaction avec un agent 
 (Osherenko et al., 2009) mais celles-ci restent envisagées indépendamment de la
 stratégie de dialogue.

Les développements de la thèse s’intègreront dans la plateforme GRETA qui repose
 sur l’architecture SAIBA, une architecture globale unifiée développée par la 
“communauté agent” pour la génération de comportements multimodaux 
 (Niewiadomski et al., 2011). Greta permet de communiquer avec l’humain en
 générant chez l’agent une large palette de comportements expressifs verbaux et
 non verbaux (Bevacqua et al., 2012). Elle peut simultanément montrer des 
expressions faciales, des gestes, des regards et des mouvements de têtes. Cette 
plateforme a notamment été intégrée dans le cadre du projet SEMAINE avec le
 développement d’une architecture temps-réel d’interaction humain-agent 
(Schröder et al., 2011) qui inclut des analyses acoustiques et vidéos, un 
système de gestion du dialogue et, du côté de la synthèse, le système Text To
 Speech OpenMary et l’agent virtuel de la plateforme GRETA. A l’instar de ce 
projet, la détection d’opinions et de sentiments envisagée dans la thèse
 interviendra en entrée des modèles d’interactions multi-modaux de la
 plateforme. La stratégie de dialogue multimodale associée à ces entrées
 relatives au contenu verbal devra être définie et intégrée dans la plateforme
 GRETA.

**Enjeux
La thèse portera sur le développement conjoint de méthodes de détection des
 opinions et des sentiments et de stratégies de dialogue humain-agent. Les 
méthodes envisagées sont des méthodes hybrides mêlant apprentissage statistique
 et règles expertes.  Pour les stratégies de dialogue, le doctorant pourra
 s’appuyer sur les travaux réalisés dans le cadre du moteur de dialogue DISCO
 (Rich et al., 2012) et du moteur développé dans le projet Semaine(Schröder et
 al., 2011). Les méthodes développées pourront également s’appuyer sur des
 analyses de corpus humain-humain ou de type Magicien d’Oz (McKeown et al.,
 2012) et un protocole d’évaluation de ces méthodes devra être mis en place. En
 particulier, pour répondre à cet objectif, la thèse devra aborder les
 problématiques suivantes:
-  la définition des types d’opinions et de sentiments pertinents à considérer
 en entrée du moteur de dialogue. Il s’agira d’aller au-delà delà de la
 distinction classique entre opinions positives et opinions négatives, peu
 pertinente dans ce contexte, en s’appuyant sur les modèles issus de la
 psycholinguistique (Martin and White, 2007);
- l’identification des marqueurs lexicaux, syntaxiques, sémantiques et 
dialogiques des opinions et des sentiments;
 - la prise en compte du contexte d’énonciation: les règles implémentées
 pourront intégrer différentes fenêtres d’analyse : la phrase, le tour de parole
 et les tours de paroles antérieurs;
-  la prise en compte des problématiques temps-réel de l’interaction : des
 stratégies de dialogues seront définies en fonction des différentes fenêtres
 d’analyse afin de proposer des stratégies d’interactions à différents niveaux 
de réactivité. Par exemple, certains mots-clés pourront être utilisés comme
 déclencheurs de backchannel en temps réels et la planification des
 comportements de l’agent pourra être ajustée au fur et à mesure de l’avancement 
de l’interaction.

**Ouverture à l’international:
Ces travaux de thèse interviennent en complémentarité des travaux réalisés sur
 les interactions non verbales dans le cadre du projet européen FP7 TARDIS
 prenant comme application les Serious games dans le cas d’un entrainement à 
l’entretien d’embauche (http://tardis.lip6.fr/presentation) et des travaux
 réalisés sur le traitement des signaux sociaux dans le cadre du réseau
 d’excellence SSPNET (http://sspnet.eu/) Une collaboration avec Candy Sidner,
 professeur au département Computer Science du Worcester Polytechnic Institute 
et experte en modèles computationnels d’intéractions verbales et non verbales et
 à l’origine du moteur de dialogue DISCO (Richet et al. 2012) sera également
 mise en place.

**Références:
E. Bevacqua, E. de Sevin, S.J. Hyniewska, C. Pelachaud (2012), A listener model:
 Introducing personality traits, Journal on Multimodal User Interfaces, special 
issue Interacting ECAs, Elisabeth André, Marc Cavazza and Catherine Pelachaud 
(Guest Editors), 6:27–38, 2012.
M. Chollet, M. Ochs and C. Pelachaud (2012), Interpersonal stance recognition
 using non-verbal signals on several time windows, Workshop Affect, Compagnon
 Artificiel, Interaction, Grenoble, November 2012, pp. 19-26
C. Clavel and G. Richard (2010). Reconnaissance acoustique des émotions, 
Systèmes d’interactions émotionnelles, C. Pelachaud,  chapitre 5, 2010
C. Clavel, G. Adda, F. Cailliau, M. Garnier-Rizet, A. Cavet, G. Chapuis, S.
 Courcinous, C. Danesi, A-L. Daquo, M. Deldossi, S. Guillemin-Lanne, M. Seizou,
 P. Suignard (2013). Spontaneous Speech and Opinion Detection: Mining Call
-centre Transcripts. In Language Resources and Evaluation, avril 2013.
M. El-Bèze, A. Jackiewicz, S. Hunston, Opinions, sentiments et jugements
 d’évaluation, Revue TAL 2010, Volume 51 Numéro 3.
J.R. Martin , P.R.R. White (2007) Language of Evaluation: Appraisal in English, 
Palgrave Macmillan, Novembre 2007
G. McKeown, M. Valstar, R. Cowie, R., M. Pantic, M. Schroder (2012) The SEMAINE
 Database: Annotated Multimodal Records of Emotionally Colored Conversations
 between a Person and a Limited Agent, IEEE Transactions on Affective Computing,
 Volume: 3  , Issue: 1, Page(s): 5- 17, Jan.-March 2012
R. Niewiadomski, S. Hyniewska, C. Pelachaud (2011), Constraint-Based Model for
 Synthesis of Multimodal Sequential Expressions of Emotions, IEEE Transactions of Affective Computing, vol. 2, no. 3, 134-146, Juillet 2011.
A. Osherenko, E. Andre, T. Vogt (2009),  Affect sensing in speech: Studying fusion of linguistic and acoustic features,  International Conference on Affective Computing and Intelligent Interaction and Workshops, 2009
C. Rich, C. L. Sidner (2012), Using Collaborative Discourse Theory to Partially Automate Dialogue Tree Authoring. IVA 2012: 327-340
M. Schröder, E. Bevacqua, R. Cowie, F. Eyben, H. Gunes, D. Heylen, M.ter Maat, G. McKeown, S. Pammi, M. Pantic, C. Pelachaud, B. Schuller, E. de Sevin, M.l Valstar, and M. Wöllmer (2011), Building Autonomous Sensitive Artificial Listeners, IEEE Transactions of Affective Computing, pp. 134-146, Octobre 2011.
P. Suignard, (2010)  NaviQuest : un outil pour naviguer dans une base de questions posées à un Agent Conversationnel, WACA, Octobre 2010

Back

Top

PhD student Learning  Pronunciation Variants in a Foreign Language (full time)   

Back

Top

Available PhD position with scholarship - Silent speech interface

GIPSA-lab, Grenoble, France

Incremental speech synthesis for a real-time silent speech interface

Context

: The design of a silent speech interface, i.e. a device allowing speech communication without the

necessity of vocalizing the sound, has recently received considerable attention from the speech research

community [1]. In the envisioned system, the speaker articulates normally but does not produce any audible

sound. Application areas are in the medical field, as an aid for larynx-cancer patients, and in the

telecommunication sector, in the form of a “silent telephone”, which could be used for confidential

communication, or in very noisy environments. In [2], we have shown that ultrasound and video imaging can be

efficiently combined to capture the articulatory movements during silent speech production; the ultrasound

transducer and the video camera are placed respectively beneath the chin and in front of the lips. At present, our

work focused mainly on the estimation of the target spectrum from the visual articulatory data (using artificial

neural network, Gaussian mixture regression and hidden Markov modeling). The other challenging issue

concerns the

estimation of acceptable prosodic patterns (i.e. the intonation of the synthetic voice) from silent

articulatory data only. To address this ill-posed problem, one solution consists of splitting the mapping process

into two consecutive steps: (1) a visual speech recognition step which estimates the most likely sequence of word

given the articulatory observations, and (2) a text-to-speech (TTS) synthesis step which generates the audio signal

from the decoded word sequence. In that case, the target prosodic pattern is derived from the linguistic structure

of the decoded sentence. The major drawback of this mapping method is that it cannot run in real-time. In fact, if

the visual speech recognition step can be done

online (i.e. words are decoded a short amount of time after they

have been pronounced), standard TTS systems need to know the entire sentence to estimate the target prosody.

This introduces a large delay between the (silent) articulation and the generation of the synthetic audio signal.

This delay prevents the communication partners from having a fluent conversation. The main goal of this PhD

project is to design a

real-time silent speech interface, in which the delay between the articulatory gesture and

the corresponding acoustic event has to be constant and as short as possible.

Goals:

The goal of this PhD project is twofold:

(1) Reducing the delay between the recognition

and the synthesis steps, by designing a new

generation of TTS system, called “

incremental

TTS system

” [3]. This system should be able to

synthesize the decoded words, with acceptable

prosody, as soon as they are provided by the

visual speech recognition system.

(2) Designing experimental paradigms in order to evaluate the system in realistic communication situations (faceto-

face, remote/telephone-like interaction, human-machine interaction). The goal is to study how a silent speaker

benefits from the acoustic feedback provided by the incremental TTS and how he/she adapts his/her own

articulation to maximize the efficiency of the communication.

Supervision:

Dr. Thomas Hueber, Dr. Gérard Bailly (CNRS/GIPSA-lab)

Duration / Salary

: 36 months (October 2013- October 2016) / ~ €1400/month minimum (net salary).

Research fields

: multimodal signal processing, machine learning, interactive systems, experimental design

Background:

Master’s or engineer’s degree in computer science, signal processing or applied mathematics.

Skills

: Good skills in mathematics (machine learning) and programming (Matlab, C, Max/MSP). Knowledge in

speech processing or computational linguistics would be appreciated.

To apply

: send your CV, transcript of records of your Master grade and a cover letter to

thomas.hueber@gipsa-lab.grenoble-inp.fr

References

:

[1] B. Denby, T. Schultz, K. Honda, T. Hueber, et al., “Silent Speech Interfaces,” Speech Communication, vol. 52, no. 4, pp.

270-287, 2010.

[2] T. Hueber, E. L. Benaroya, G. Chollet, et al., “Development of a Silent Speech Interface Driven by Ultrasound and

Optical Images of the Tongue and Lips”, Speech Communication, vol. 52, no. 4, pp. 288-300, 2010.

[3] Buschmeier H, Baumann T, Dosch B, Schlangen D, Kopp S. “Combining Incremental Language Generation and

Incremental Speech Synthesis for Adaptive Information Presentation”, in proc of the 13th Sigdial meeting, pp, 295-303,

2012.

Back

Top

ance'Offre de post-doctorat dans le cadre du projet ANR DIADEMS (Description, Indexation, Accès aux Documents Ethnomusicologiques et Sonores).

- Sujet de post-doctorat : identification / classification instrumentale

Durée : 12 mois

Salaire : environ 2000 €/mois

Date de début souhaitée : septembre 2013

La reconnaissance automatique d'instrument et la classification par famille d'instruments est un domaine de recherche actif du MIR (Music Information Retrieval) [Hei09] [Kit07] [Her06] [Ess06]. Les principales techniques reposent sur des méthodes statistiques utilisant des paramètres audio de type MFCC. Nous souhaitons ici tracer une voie nouvelle, permettant de faire le lien entre le traitement de la parole et le traitement de la musique, en considérant l'interprétation musicale comme une phrase, et l'instrument ou l'instrumentiste comme un locuteur.

Ce travail s'effectuera en parallèle d'une thèse en cours sur la caractérisation et l'identification de la voix chantée. Au cours de cette thèse, nous avons proposé une méthode permettant d'identifier les segments contenant de la voix chantée dans des enregistrements polyphoniques (e.g. musique 'pop'). L'objet actuel d'étude est de déterminer quels sont les paramètres du signal les plus pertinents pour caractériser différents styles de chant.

Une des pistes que nous souhaitons poursuivre sera d'identifier l'instrument en suivant le vibrato, de manière similaire à ce qui est proposé pour la voix chantée. En insistant sur la dimension temporelle plutôt que spectrale, nous pourrons aussi observer comment s'enchainent les respirations, les attaques sonores ou les changements timbraux utilisés par le musicien. Ce travail exploratoire nécessitera dans un premier temps d'effectuer des expérimentations sur des bases de données simples (telles que [Fri97] et [Got03]) afin de valider notre approche avant d'appliquer nos algorithmes aux données du projet DIADEMS.

- Références :

[Hei09] Heittola, T., Klapuri, A., Virtanen, T., 'Musical Instrument Recognition in Polyphonic Audio Using Source-Filter Model for Sound Separation,' in Proc. 10th Int. Society for Music Information Retrieval Conf. (ISMIR 2009), Kobe, Japan, 2009.

[Kit07] Tetsuro Kitahara, Masataka Goto, Kazunori Komatani, Tetsuya Ogata, and Hiroshi G. Okuno: 'Instrument Identification in Polyphonic Music: Feature Weighting to Minimize Influence of Sound Overlaps', EURASIP Journal on Advances in Signal Processing, Special Issue on Music Information Retrieval based on Signal Processing, Vol.2007, No.51979, pp.1--15, 2007.

[Her06] P. Herrera-Boyer, A. Klapuri, and M. Davy. Automatic classification of pitched musical instrument sounds. Signal Processing Methods for Music Transcription, pages 163–200. Springer, 2006.

[Ess06] S. Essid, G. Richard, and David.B. Instrument recognition in polyphonic music based on automatic taxonomies. IEEE Transactions on Audio, Speech & Language Processing, 14(1):68–80, 2006.

[Fri97] L. Fritts, “Musical Instrument Samples,” Univ. Iowa Electronic Music Studios, 1997–. [Online]. Available: http://theremin.music.uiowa.edu/MIS.html

[Got03] Goto M, Hashiguchi H, Nishimura T, Oka R. RWC music database: Music genre database and musical instrument sound database. ISMIR. 2003:229–230.

---------

Description du projet DIADEMS (Partenaires : LaBRI, IRIT, LESC, Parisson, LIMSI, MNHN, LAM-IJLRA)

Le Laboratoire d'Ethnologie et de Sociologie Comparative (LESC) comprenant le Centre de Recherche en Ethnomusicologie (CREM) et le centre d'Enseignement et de Recherche en Ethnologie Amérindienne (EREA) ainsi que le Laboratoire d'Eco-anthropologie du Muséum National d'Histoire Naturelle (MNHN) sont confrontés à la nécessité d'indexer les fonds sonores qu'ils gèrent et de faire un repérage des contenus, travail long, fastidieux et coûteux.

Lors de l'Ecole d'Été interdisciplinaire Sciences et Voix 2010 organisée par le CNRS, une convergence d'intérêts s'est dégagée entre les acousticiens, les ethnomusicologues et les informaticiens : il existe aujourd'hui des outils d'analyse avancés du son développés par les spécialistes en indexation qui permettent de faciliter le repérage, l'accès et l'indexation des contenus.

Le contexte du projet est l'indexation et l'amélioration de l'accès aux fonds d'archives sonores du LESC : le fonds du CREM et celui d'ethnolinguistique de l'EREA (« chanté-parlé » Maya, ainsi que celui du MNHN (musique traditionnelle africaine). Il s'inscrit dans la continuité d'une réflexion entreprise en 2007 pour l'accès aux données sonores de la Recherche : aucune application n'existant en « open source » sur le marché, le CREM-LESC, le LAM et la Phonothèque de la MMSH d'Aix-en Provence ont étudié la conception d'un outil innovant et collaboratif qui répond à des besoins « métier » liés à la temporalité du document, tout en étant adapté à des exigences du secteur de la recherche. Avec le soutien financier du Très Grand Equipement (TGE) ADONIS du CNRS et du Ministère de la Culture, la plateforme Telemeta développée par la société PARISSON a été mise en ligne en mai 2011 : http://archives.crem-cnrs.fr . Sur cette plateforme, des outils d'analyse élémentaires de traitement de signal sont d'ores et déjà disponibles.

Cependant, il est nécessaire de disposer d'un ensemble d'outils avancés et innovants pour une aide à l'indexation automatique ou semi-automatique de ces données sonores, issues d'enregistrements parfois longs, au contenu très hétérogène et d'une qualité variée. L'objectif du projet DIADEMS est de fournir certains des outils, de les intégrer dans Telemeta, en répondant aux besoins des usagers. Il s'en suit une complémentarité des objectifs scientifiques des différents partenaires : Les fournisseurs de technologies, l'IRIT, le LIMSI, le LaBRI et le LAM auront à :

- Fournir des technologies existantes telles que la détection de parole, de musique, la structuration en locuteurs. Ces outils visent à extraire des segments homogènes d'intérêt pour l'usager. Ces systèmes auront à faire face à la diversité des bases qu'il est proposé d'étudier dans ce projet ; leur hétérogénéité est liée aux conditions d'enregistrement, au genre et à la nature des documents, à leur origine géographique. Il faudra adapter ces systèmes dits « état de l'art » aux besoins des usagers.

- Proposer des outils innovants d'exploration du contenu de segments homogènes. Les travaux sur l'opposition voix parlée-déclamée-chantée, le chant, les tours de chant, la recherche de similarité musicale ne sont pas matures. Un véritable travail de recherche reste à faire et avoir à sa disposition des musicologues et des ethnomusicologues est un atout positif. Les ethnomusicologues, ethnolinguistes, acousticiens spécialistes de la voix et les documentalistes spécialisés vont jouer un rôle important dans le projet en tant que futurs utilisateurs des outils d'indexation : Les documentalistes doivent s'approprier les outils et apporter leur expérience afin d'adapter ces outils à leur besoin en indexation.

Un échange important doit se réaliser entre celui qui fournit l'outil, celui qui l'intègre et celui qui l'utilise. L'effort doit être porté sur la visualisation des résultats avec pour fin une aide forte à l'indexation en la rendant de fait semi-automatique Pour l'ethnomusicologue et le musicologue, l'objectif va au-delà de l'indexation. Il s'agit au travers d'aller et retour entre lui et les concepteurs de technologies de cibler les outils pertinents d'extraction d'information.

Jean-Luc Rouas LABRI 351, cours de la Libération 33405 Talence cedex France (+33) 5 40 00 35 08

Back

Top

Reconnaissance du locuteur en milieu bruité

Nous avons atteint ces dernières années de très bonnes performances en reconnaissance du locuteur. Et ce, malgré la présence de la variabilité session. En effet, le variabilité session est prise en compte lors du scoring en utilisant une matrice de covariance modélisant cette dernière. Ce processus est effectué dans l'espace des i-vectors [1]. Le concept des i-vectors est devenu un standard en reconnaissance du locuteur.

Dans la dernière évaluation internationale NIST 2012, nous avons été confrontés à une nouvelle difficulté qui est le bruit additif [2], c'est à dire le bruit ambiant. La recherche pour réduire l'impact du bruit dans les systèmes de reconnaissance du locuteur est motivée en grande partie par le besoin d'appliquer les technologies de reconnaissance du locuteur sur des appareils portables ou sur l'Internet. Alors que les technologie promet un niveau supplémentaire de sécurité biométrique pour protéger l'utilisateur, la mise en œuvre pratique de ces systèmes doit faire face à de nombreux défis. Un des plus importants défis à surmonter est le bruit environnemental. En raison de la mobilité de ces systèmes, les sources de bruit peuvent être très variables dans le temps et potentiellement inconnus.

Nous proposons de travailler dans ce cadre : proposer des stratégies permettant de compenser l'effet du bruit additif, ces stratégies peuvent intervenir à différents niveaux du processus de reconnaissance: au niveau du signal, au niveau des modèles acoustiques, au niveau des i-vectors et au niveau du scoring....) .

• Débruitage des signaux

• Effet du bruit sur la VAD (Voice activity detection)

• Bruitage des modèles

• Intégration des caractéristiques statistiques du bruit dans la phase du scoring

Dans une deuxième partie du travail, nous proposons de nous mettre dans les meilleures conditions pour que le système soit le plus robuste au bruit. Par exemple, le choix de l'énoncé à prononcer par le locuteur peut avoir de l'influence sur les performances du système [3]. Faut a t-il avoir avoir le même énoncé pour tous les locuteurs, ou au contraire chaque locuteur se distingue des autres locuteur sur un ensemble bien précis d'unités acoustiques. Dans ce dernier cas, il faut trouver une stratégie, qui permet de déterminer l'ensemble des unités acoustiques qui différencient le plus possible un locuteur (des autres locuteurs). D'autres stratégies de robustesse au bruit doivent être proposées et étudiées dans le cadre de cette thèse. Une des pistes à explorer est l'utilisation de la théorie des caractéristiques manquantes (missing-feature theory), qui a été utilisée dans le domaine du traitement de la parole [4][5][6].

Les systèmes de reconnaissance du locuteur de l'état de l'art sont fondamentalement basés sur l'utilisation de l'UBM (Universal Backgroud Model), il s'agit d'un modèle trop simple pour le traitement et la modélisation de la parole. Dans le cas de la reconnaissance en milieu bruité, la tâche devient plus complexe, il est donc légitime de se reposer la question sur l'adéquation de ce modèle pour cette tâche. Nous proposons d'adapter une approche utilisant des HMM (ou autre modèle) à cette tâche tout en profitant des avancées récemment proposées ( Factor analysis, I-vectors, …).

[1] Bousquet Pierre-Michel, Matrouf Driss and Bonastre Jean-François, «Intersession compensation and scoring methods in the i-vectors space for speaker recognition » Interspeech 2011, Florence.

[2] Miranti Indar Mandasari, Mitchell McLaren and David A. van Leeuwen, « The Effect of noise on modern automatic speaker recognition systems » , ICASSP 2012.

[3] Anthony Larcher, Pierre-Michel Bousquet, Kong-Aik Lee, Driss Matrouf, Haizhou Li, Jean-François Bonastre, « I-vectors in the context of phonetically-constrained short utterances for speaker verification. » ICASSP 2012: 4773-4776.

[4] M.P. Cooke, P.G. Green, L. Josifovski, and A. Vizinho, « Robust ASR with

unreliable data and minimal assumptions, » in Proc., Robust’99, 1999

[5] M.P. Cooke, P.G. Green, L. Josifovski, and A. Vizinho, « Robust Automatic Speech Recognition with missing and unreliable acoustic data, » Speech Communication,, 2000.

[6] B. Raj, M.L. Seltzer, and R.M. Stern, « Reconstruction of missing features for robust speech recognition, » Speech Communication, 2004.

Back

Top

Le labex Brain and Language Research Institute (BLRI) financera à la rentrée prochaine deux sujets 

de thèse dont l'un des sujets pourra être celui proposé ci-dessous (en fonction des candidatures

 retenues).

Calendrier :. date limite des candidatures : 10 juin. 

auditions des candidats retenus : 24 juin

Bourse : 1 684.93€ brut mensuel (1 368€ nets)

Dossier des candidats :. CV détaillé. Notes. Motivation et/ou projet scientifique correspondant au sujet

Contact scientifique : Corinne Fredouille et Christine Meunier

Contact administratif : nadera.bureau@blri.fr

Description du sujet : 

Titre : Détection de zones de déviance dans la parole pathologique : apport du traitement automatique face à l'expertise humaine

Superviseurs : Corinne Fredouille, Christine Meunier

Laboratoire d'accueil : Laboratoire Informatique d'Avignon (collaboration avec le Laboratoire Parole et Langage –Aix-en-Provence)

Discipline et école doctorale : Informatique, école doctorale ED536 de l'Université d'Avignon

Calendrier :

. date limite des candidatures : 10 juin

. auditions des candidats retenus : 24 juin

Bourse : 1 684.93€ brut mensuel (1 368€ nets)

Description scientifique

Si la définition de l'étendue de la variabilité en parole normale est une question fondamentale pour les théories

linguistiques actuelles, une façon d'aborder ses limites est d'essayer de déterminer ses frontières par le biais de la

variation pathologique. Comme le soulignent Duffy et Kent, 2001 « Science often takes advantages of nature’s

accidents to learn the principles of a process ». Sur ce principe, la connaissance de la parole pathologique s'appuyant sur

la compréhension des phénomènes d'altération « observables » dans la production de parole de patients atteints de

troubles de la parole devient une nécessité.

La parole dysarthrique correspond à une altération de la commande motrice d’origine centrale ou périphérique des

gestes de la parole. Des variations importantes existent dans la parole dysarthrique en relation avec un déficit de

l’exécution temporo-spatiale des mouvements de la parole et qui peuvent affecter différents niveaux de production

(respiratoire, laryngé et supralaryngé). La grande majorité des travaux ayant porté sur l'étude de la parole dysarthrique

repose sur des analyses perceptives. La raison principale tient dans le fait qu'un patient dysarthrique est dysarthrique

parce qu'il « s'entend/a l'air » dysarthrique. Les travaux les plus connus au niveau international sont ceux de Darley et

al., 1975. Ils ont conduit à l'élaboration d'une organisation des dysarthries en 6 classes (complétée par deux classes

supplémentaires par Duffy, 1995) sur la base de clusters physiopathologiques définis à partir de la concomitance de

caractéristiques les plus déviantes perçues par un jury d'écoute. L’hypothèse sous-jacente à la construction de ces

clusters est qu’un ensemble de paramètres perturbés simultanément, mis en relation avec l’atteinte neurologique,

reflèterait un processus physiopathologique particulier. Si cette classification reste d'actualité encore aujourd'hui pour

évaluer notamment la parole dysarthrique en pratique clinique, elle reste néanmoins sujette à controverses pour deux

raisons principales : la subjectivité des évaluations perceptives d'une manière générale et la difficulté pour un être

humain, même expérimenté, à distinguer et à juger perceptivement les multiples dimensions à prendre en compte dans

l'évaluation de la parole dysarthrique. En conséquence, différents travaux ont été menés à partir des années 80 à

aujourd'hui dans l'objectif de combiner ces analyses perceptives à des méthodes plus objectives et quantitatives telles

que les analyses instrumentales basées sur des mesures acoustiques ou physiologiques (revue de la littérature dans Kay,

2012). Si les analyses instrumentales peuvent s'appuyer sur des traitements semi- voire entièrement automatiques,

l'analyse acoustique fine nécessaire pour comprendre les phénomènes déviants inhérents à la dysarthrie dans le signal de

parole demeure encore très coûteuse en temps par un expert humain. Dans ce contexte, une grande part des études

présentes dans la littérature repose soit sur un nombre très restreint de patients ou sur une pathologie bien ciblée.

Pourtant, la grande variabilité des phénomènes déviants observés dans la parole dysarthrique en fonction de la

pathologie du patient, de l'avancement de la maladie ou de la sévérité de la dysarthrie requiert d'analyser une large

population de patients.

L'objectif de cette thèse est d'étudier comment les outils du traitement automatique de la parole pourraient permettre de

traiter de larges populations de patients dysarthriques et de focaliser l'attention des experts humains sur des zones de

déviance bien identifiées du signal en vue d'analyses plus fines. Ces travaux reposeront notamment sur le système de

transcription automatique du LIA et ses activités de recherche autour des mesures de qualité des transcriptions

(Lecouteux, 2008 et Senay, 2011). La granularité de la détection des zones de déviance – ici potentiellement le mot ou

la séquence de mots – sera dans un second temps affiner par des outils travaillant à des niveaux inférieurs allant

jusqu'au phonème (Fredouille, 2011).

Ces travaux devront tenter de répondre à différentes questions :

• Face à la variabilité des phénomènes de déviance observés dans la parole dysarthrique et répertoriés dans la

littérature, quels sont ceux qu'un système de détection automatique est capable de déceler ?

• Est ce qu'un système automatique est capable de mettre en évidence les mêmes phénomènes de déviance qu'un

expert humain lors d'une évaluation perceptive ?

• Les déviances détectées par le système automatique sont-elles pertinentes pour les phonéticiens ?

• Est-il possible de mettre en relation les déviances détectées avec la physiopathologie du patient (ex : indices

hypokinéthiques pour la maladie de Parkinson, indices paralytiques pour la SLA, …) ?

Les travaux autour du système de transcription automatique du LIA devraient également ouvrir des perspectives sur la

mise en place d'un système de mesures objectives de l'intelligibilité des patients dysarthriques.

Ces travaux de thèse seront réalisés dans le cadre d'une collaboration étroite entre le LIA (Corinne Fredouille) pour son

expertise autour des systèmes automatiques, le LPL (Christine Meunier et Alain Ghio) pour son expertise sur les

analyses acoustico-phonétiques et les évaluations perceptives, les hôpitaux de la Timone (Dr Danièle Robert) et des

Pays d'Aix (François Viallet) pour leur expertise clinique. Ils seront basés sur le corpus de patients dysarthriques

élaborés dans le cadre du projet ANR DesPhoAPady (2009-2012 – Fougeron, 2010). Ce corpus présente un large panel

de patients souffrant de différentes pathologies (maladie de Parkinson, Sclérose Latérale Amiotrophique, syndrôme

cérébelleux) et différents niveaux de sévérité de dysarthrie.

Références :

J. R. Duffy, R. D. Kent, « Darley's contributions to the understanding, differential diagnosis, and scientific study of the

dysarthrias », Aphasiology 15(3):275 – 289, 2001.

F. L. Darley, A. E. Aronson, J. R. Brown, « Motor Speech Disorders », Philadelphia: W.B. Saunders, 1975.

J. R. Duffy, « Motor speech disorders : substrates, differential diagnosis and management », Motsby- Yearbook, St

Louis, 1e édition, 1995.

T. S. Kay, « Spectral analysis of stop consonants in individuals with dysarthria secondary to stroke », PhD thesis,

Department of Communication Sciences and Disorders, Faculty of the Louisiana State University and Agricultural and

Mechanical College, USA, 2012.

B. Lecouteux, « Reconnaissance automatique de la parole guidée par des transcriptions a priori », Thèse de doctorat,

Université d'Avignon et des Pays Vaucluse, 2008.

G. Senay, « Approches semi-automatiques pour la recherche d’information dans les documents audio », Thèse de

doctorat, Université d'Avignon et des Pays Vaucluse, 2011.

C. Fredouille, G. Pouchoulin, « Automatic detection of abnormal zones in pathological speech », International Congress

of Phonetic Sciences (ICPHs'11), Hong Kong, 17-21 Août 2011.

C. Fougeron et al., « Developing an acoustic-phonetic characterization of dysarthric speech in French », LREC'10

(international conference on Language Resources and Evaluation), Malte, Mai 2010.

Dossier des candidats :

. CV détaillé

. Notes

. Motivation et/ou projet scientifique correspondant au sujet

Contact scientifique : Corinne Fredouille et Christine Meunier

Contact administratif : nadera.bureau@blri.fr

Title: Detection of deviant zones in pathological speech : contribution of the automatic speech processing against the

Human expertise

Supervisor : Corinne Fredouille, Christine Meunier

Host laboratory : Laboratoire Informatique d'Avignon (collaboration with the Laboratoire Parole et Langage – Aix-en-Provence)

Field and doctoral school : computer sciences, doctoral school ED536 of the University of Avignon

Date :

. deadline for the application : 10th of june

. auditions of chosen candidates : 24th of june

Grant : 1 684.93€ montly gross (1 368€ net)

Scientific description

If the definition of the variability range in normal speech is a key issue for the current linguistic theories, a way of

dealing with its limits is to attempt to determine its frontiers through pathological variation. As reported by Duffy and

Kent, 2001 « Science often takes advantages of nature’s accidents to learn the principles of a process ». Based on this,

the knowledge of the pathological speech, based on the understanding of alteration phenomena, « observable » on the

speech production of patients suffering of speech disorders becomes a necessity.

Dysarthria is a group of speech disorders resulting from neurological impairments of speech motor control. Substantial

variations occur in dysarthric speech due to a deficit in the spatio-temporal execution of speech movements that affects

different levels of speech production (respiratory, laryngeal and supralaryngeal). The vast majority of research work

dedicated to the study of the dysarthric speech relies on perceptual analyses. The main reason is that a dysarthric patient

is dysarthric because he/she sounds dysarthric. The most known study, at international level, was done by Darley et al.,

1975. This work leads to organize dysarthria into 6 classes (completed with 2 additional classes by Duffy, 1995) on the

basis of physiopathological clusters defined from the co-occurrences of the most deviant features perceived by a

perceptual jury. The hypothesis underlying the building of these clusters is that a set of simultaneously disturbed

features, connected with the neurological injuries, should reflect a typical physiopathological process.

If this classification is still used nowadays to evaluate dysarthric speech in clinical practice notably, it remains

controversal for a couple of reasons : the subjectivity of perceptual evaluation and the difficulty for a Human being,

even with a high expertise, of distinguishing and assessing perceptually the multiple dimensions to take into account

when dealing with dysarthric speech. Consequently, different research work has been conducted since the 1980s until

now which aims at combining these perceptual analyses with more objective and quantitative approaches such as the

instrumental analyses based on acoustic or physiological measure (review of the literature can be found in Kay, 2012).

Contrary to the instrumental analyses which can rely on some semi- or full-automatic process, in-depth acoustic

analysis of speech necessary to understand the deviant phenomena related to dysarthria still remains very timeconsuming

for a Human expert. Based on this, a signifiant proportion of studies in the literature are conducted on a

limited number of patients or on a focused pathology. However, the large variability of deviant phenomena observed in

dysarthric speech according to the patient’s pathology, the stage of the disease, or the dysarthria severity require the

analysis of a large patient population.

The aim of this thesis is to study how the automatic speech processing tools could permit to treat large populations of

dysarthric patients and to focus Human experts’ attention on speech zones well identified as deviant for further in-depth

analyses. This work will rely on the automatic speech transcription developed at the LIA and its activities on the quality

measure of transcriptions (Lecouteux, 2008 et Senay, 2011). The granularity of the deviant zone detection – here the

word or set of words – will be refined, in a second step, by applying existing detection tools working at lower levels like

the phoneme (Fredouille, 2011).

This work will attempt to answer the following key issues :

• Given the variability of deviant phenomena observed on dysarthric speech and reported in literature, which

ones is an automatic detection system able to capture ?

• Is an automatic system able to highlight the same deviant phenomena as a Human expert will detect

perceptually ?

• Are deviant speech zones detected by an automatic system relevant for a phonetician ?

• Does a correlation between the type of deviant phenomena detected and the patient’s physiopathology

exist (e.g : hypokinetic feature for the Parkinson disease, paralytic features for ALS, …) ?

Studies relating to the automatic speech transcription should open up new perspectives on the implementation of an

objective system dedicated to the evaluation of the dysarthric patient’s intelligibility.

This thesis work will be carried out within a close collaboration between the LIA (Corinne Fredouille) for her expertise

on the automatic system dedicated to speech processing, the LPL (Christine Meunier and Alain Ghio) for their expertise

on acoustic-phonetic analyses and perceptual evaluations, both the hospitals of La Timone (Dr Danièle Robert) and desPays d’Aix (Pr. François Viallet) for their clinical expertise. It will be based on the dysarthric patient corpus designed

for the ANR DesPhoAPady project (2009-2012 – Fougeron, 2010). This corpus includes a large population of patients

suffering from various pathologies (Parkinson disease, ALS, cerebelar syndrom, …) and different levels of dysarthria

severity.

Bibliography :

J. R. Duffy, R. D. Kent, « Darley's contributions to the understanding, differential diagnosis, and scientific study of the

dysarthrias », Aphasiology 15(3):275 – 289, 2001.

F. L. Darley, A. E. Aronson, J. R. Brown, « Motor Speech Disorders », Philadelphia: W.B. Saunders, 1975.

J. R. Duffy, « Motor speech disorders : substrates, differential diagnosis and management », Motsby- Yearbook, St

Louis, 1e édition, 1995.

T. S. Kay, « Spectral analysis of stop consonants in individuals with dysarthria secondary to stroke », PhD thesis,

Department of Communication Sciences and Disorders, Faculty of the Louisiana State University and Agricultural and

Mechanical College, USA, 2012.

B. Lecouteux, « Reconnaissance automatique de la parole guidée par des transcriptions a priori », Thèse de doctorat,

Université d'Avignon et des Pays Vaucluse, 2008.

G. Senay, « Approches semi-automatiques pour la recherche d’information dans les documents audio », Thèse de

doctorat, Université d'Avignon et des Pays Vaucluse, 2011.

C. Fredouille, G. Pouchoulin, « Automatic detection of abnormal zones in pathological speech », International Congress

of Phonetic Sciences (ICPHs'11), Hong Kong, 17-21 Août 2011.

C. Fougeron et al., « Developing an acoustic-phonetic characterization of dysarthric speech in French », LREC'10

(international conference on Language Resources and Evaluation), Malte, Mai 2010.

Candidate application form :

. detailed CV

. marks

. Motivation and/or scientific project related to the topic

Scientific Contact : Corinne Fredouille et Christine Meunier

Administrative Contact : nadera.bureau@blri.fr

Back

Top

At the Centre for Next Generation Localisation (CNGL) in Dublin, Ireland, we have a number of internships available covering a wide range of topics in Natural Language Processing and Machine Translation based at our Dublin City University site.  The internships are available for both basic research and more applied research projects (including development-focused work).
Candidates are required to be registered as MSc or PhD students (by research) in their home universities while carrying out their internship in Dublin and need to provide written confirmation of this from their home institute. Please find the internship advertisement attached.
Details of a number of specific internships can be found at: http://www.cngl.ie/outreach/graduate-programme/postgradinternships/
Closing date: 31st May 2013
For any informal enquiries please contact: CNGL Education and Outreach
Dr. Cara Greene, CNGL, DCU Phone: +353 (0)1 7006704 E-mail: cgreene (AT) computing.dcu.ie
Web: http://www.cngl.ie
Application Procedure:
For formal applications, please download an application form from the link below and send it to cgreene (AT) computing.dcu.ie by Friday 31st May 2013. http://www.cngl.ie/outreach/graduate-programme/postgradinternships/

Back

Top

vacant position as a

Speech Technology Researcher/Developer (main focus ASR)

in our research and development group in Dresden. We are a young
company located in Dresden dealing with speech technology products and
applicatuions, signal processing solutions and pattern recognition
techniques. Although we develop a new group of products our work is
closely related to scientific topics and we collaborate with several
german and international universities.

Job descriptions (in German) may be found in

http://www.linguwerk.de/sites/default/files
/1305_Stellenangebot_Linguwerk_0.pdf

linguwerk.de/sites/default/files/1305_Stellenangebot_Linguwerk_0.pdf

Unfortunately our webpage is currently under construction and only
available in German at the moment. Nevertheless English speaking
applicants are very welcome, all our group members speak English
fluently.

Our team consists of speech technology scientists and engineers,
education sciense specialists, creatives, and people who love
scientific challenges.

If you are interested to join and enrich our team, don't hesitate to
contact as and ask for further information. Also contact us for
information in English.

Note: The employer for this job will be the University in Dresden, but
you are a member of our joined research group within our Lab.

Dr.-Ing. Rico Petrick
CEO / Geschäftsleitung
Linguwerk GmbH          Office: +49 351 6533-3807
Schnorrstraße 70        Fax:    +49 351 6533-6965
01069 Dresden           Mobile: +49 174 943 8817
Germany                 E-Mail: rico.petrick@linguwerk.de
                        Web:    www.linguwerk.de

Back

Top

The CUI/University of Geneva seeks a qualified candidate for one


           Specialist in speech processing


for a period of 6 months (class 13; 5988.-). Depending on hiring rate (100% to 50%), the contract will be extended up to 12 months.


In the context of a newly granted Swiss National Science Foundation Agora project, the CUI will collaborate with the Phonetischen Laboratorium of Universität Zurich (UZH) on 'Swiss VoiceApp - Your voice. Your identity.” The project will combine speech recognition and processing as well as linguistic knowledge to provide a smartphone application for Swiss German dialect recognition and some information on the user’s voice.


The main tasks of successful candidate will be the development of speech recognition engine based on speech corpora of dialectal variants. Knowledge of speech software, namely HTK and Praat, are highly recommended.


Applicant should possess a master’s degree or an engineer diploma in computational linguistics, or in computer science but with strong knowledge of speech processing. Given the focus on Swiss German dialects, preference will be given to speaker of German.


The start date is expected to be during autumn/winter 2013; the position will remain open until filled.


Contact

Please send (1) your CV, (2) a copy of your degree(s), (3) a short letter explaining that your skills and your interests fit in the project, to Jean-Philippe Goldman, before July 31th 2013, preferably electronically or by mail at the address below:



Jean-Philippe Goldman

CUI / UNI-Rondeau
CH-1227 Carouge, Switzerland
Jean-Philippe.Goldman@unige.ch

========================================================================000


Le CUI / Université de Genève cherche à engager un/une


           Spécialiste en traitement de la parole


pour un emploi temporaire d’une durée de 6 mois, (classe 13 ; 5988.-).Selon le taux d’engagement (de 100% à 50%) le contrat peut être étendu à 12 mois.


Dans le contexte d’un projet récemment financé par le Fonds National de la Recherche, le CUI collaborera avec le Phonetisches Laboratorium de l’Université Zurich (UZH) sur le projet 'Swiss VoiceApp - Your voice. Your identity.” Le projet combinera des techniques de reconnaissance de la parole et des connaissances linguistiques pour produire une application pour smartphone visant à l’identification de dialectes suisses alémaniques et de fournir à l’utilisateur des informations sur sa propre voix.


La tâche principale du candidat sélectionné sera le développement d’un système de reconnaissance de la parole basé sur des corpus de variantes dialectales. La connaissance de logiciels de parole tels que HTK est essentielle.


Le candidat devra posséder un MA ou un diplôme d’ingénieur en linguistique computationnelle, ou en informatique mais avec de bonnes connaissances en traitement de la parole. Étant donné l’orientation germanophone du projet, une préférence sera donnée à un locuteur de l’allemand. Le contrat pourra commencer en automne ou hiver 2013.


Contact : Veuillez envoyer  (1) votre CV, (2) une copie de vos diplômes, (3) une lettre sur vos compétences et intérêts en rapport avec le projet,  à Jean-Philippe Goldman, avant le 31 juillet 2013, de préférence par voie électronique, ou par courrier à l’adresse suivante:


Jean-Philippe Goldman

CUI / UNI-Rondeau
CH-1227 Carouge, Switzerland
Jean-Philippe.Goldman@unige.ch

Back

Top