Search

Colloque avec actes et comité de lecture. internationale.; International audience; We present a novel noise compensation architecture which makes no assumptions on how the noise sources alter the speech data and which do not rely on clean speech models. Rather, this new architecture makes the (realistic) assumption that speech databases recorded under different background noise conditions are available. Its main principle is to process individually each database and to construct a parametric representation which describes the variation of acoustic models w.r.t. noise models. This representation is then used during recognition to estimate the acoustic models in the new environment. We evaluate the performance of this new compensation scheme on a connected digits recognition task and show that it can perform significantly better than multi-conditions training, which is the most widely used technique in these kind of scenarios.

In this thesis we focus on four principle components of a speech recognition system: acoustic modeling, language modeling, speech feature extraction and noise compensation. We propose novel modeling approaches for acoustic and linguistic modeling within the Bayesian networks formalism. Bayesian networks are a subset of probabilistic graphical models that include the most widely used probability models in speech recognition. Therefore rethinking the modeling problems in this formalism provides new perspectives that were not considered previously. Besides novel modeling approaches we also address new speech feature extraction schemes. Our main motivation in this direction is to seek for robust features that are not bound to be used in classical hidden Markov modeling (HMM) approach. Finally, we address the robustness problem for varying application conditions and propose a novel supervised compensation scheme., Dans cette thèse nous élaborons quatre composantes fondamentales d'un système de reconnaissance automatique de la parole : la modélisation acoustique, la modélisation du langage, la paramétrisation du signal acoustique et la compensation du bruit. Nous proposons des techniques nouvelles dans chacun de ces domaines, et nous apportons des perspectives novatrices. Nous traitons les problèmes de modélisation acoustique et modélisation du langage avec un outil statistique puissant : les modèles probabilistes graphiques. Ce formalisme généralise la plupart des techniques probabilistes utilisées dans le traitement de la parole. La reformulation des modules de modélisation dans ce formalisme, nous ouvre de nouvelles perspectives inexploitées auparavant. En plus des nouvelles approches pour la modélisation, nous proposons également de nouvelles stratégies pour l'extraction des paramètres acoustiques. Notre motivation principale dans ce domaine est de chercher des paramètres robustes qui ne sont pas liés à la modélisation par des HMMs. Nous abordons aussi le problème de robustesse au bruit par adaptation des modèles acoustiques et nous proposons une nouvelle méthode de compensation prédictive supervisée.

Colloque avec actes et comité de lecture. internationale.; International audience; In this paper we propose a new approach to language modeling based on dynamic Bayesian networks. The principle idea of our approach is to find the dependence relations between variables that represent different linguistic units (word, class, concept, ...) that constitutes a language model. In the context of this paper the linguistic units that we consider are syntactic classes and words. Our approach should not be considered as a model combination technique. Rather, it is an original and coherent methodology that processes words and classes in the same model. We attempt to identify and model the dependence of words and classes on their linguistic context. Our ultimate goal is to devise an automatic mechanism that extracts the best dependence relations between a word and its context, i.e., lexical and syntactic. Preliminary results are very encouraging, in particular the model in which a word depends not only on previous word but also on syntactic classes of two previous words. This model outperforms the bi-gram model.

Colloque avec actes et comité de lecture. internationale.; International audience; In this paper we propose a new approach to language modeling based on dynamic Bayesian networks. The principle idea of our approach is to find the dependence relations between variables that represent different linguistic units (word, class, concept, ...) that constitutes a language model. In the context of this paper the linguistic units that we consider are syntactic classes and words. Our approach should not be considered as a model combination technique. Rather, it is an original and coherent methodology that processes words and classes in the same model. We attempt to identify and model the dependence of words and classes on their linguistic context. Our ultimate goal is to devise an automatic mechanism that extracts the best dependence relations between a word and its context, i.e., lexical and syntactic. Preliminary results are very encouraging, in particular the model in which a word depends not only on previous word but also on syntactic classes of two previous words. This model outperforms the bi-gram model.

Colloque avec actes et comité de lecture. nationale.; National audience; We present a new predictive compensation scheme which makes no assumption on how the noise sources alter the speech data and which do not rely on clean speech models. Rather, this new scheme makes the (realistic) assumption that speech databases recorded under different background noise conditions are available. The philosophy of this scheme is to process these databases in order to build a "tool" which will allow it to handle new noise conditions in a robust way. We evaluate the performances of this new compensation scheme on a connected digits recognition task and show that it can perform significantly better than multi-conditions training, which is the most widely used techniques in these kind of scenarios.

Colloque avec actes et comité de lecture. internationale.; International audience; We present a new continuous automatic speech recognition system where no a priori assumptions on the dependencies between the observed and the hidden speech processes are made. Rather, dependencies are learned form data using the Bayesian networks formalism. This approach guaranties to improve modelling fidelity as compared to HMMs. Furthermore, our approach is technically very attractive because all the computational effort is made in the training phase.

Colloque avec actes et comité de lecture. internationale.; International audience; State-of-the-art automatic speech recognition (ASR) systems are based on probabilistic modelling of the speech signal using Hidden Markov Models. The limitations of these systems under real life conditions arose a question about the robustness of the underlying acoustic modelling methodology. The scope of my thesis is to explore the formalism of Probabilistic Graphical Models, particularly Dynamic Bayesian Networks, from a theoretical and practical point of view, with the aim of developing reliable models of speech and of developing robust ASR systems.

Colloque avec actes et comité de lecture. nationale.; National audience; Nous présentons une méthodologie pour modéliser la parole dans laquelle nous ne faisons aucune hypothèse à priori sur les dépendances entre les variables cachées et observables. Plutôt, nous donnons aux données une liberté complète pour dicter les dépendances appropriées. Cette approche a l'avantage de garantir que le modèle résultant représente la parole avec une plus grande fidélité que les HMM. En outre, un contrôle est donné à l'utilisateur pour faire un compromis entre la fidélité et la complexité du modèle. Nous évaluons le potentiel de notre approche sur une tâche de reconnaissance de chiffres connectés. || We present a speech modeling methodology where no a priori assumption is made on the dependencies between the observed and the hidden speech processes. Rather, dependencies are learned from data. This methodology guarantees improvement in modeling fid

oz GÜRÜLTÜLÜ KONUŞMANIN KALMAN SÜZGECİ İLE İYİLEŞTİRİLMESİ Deviren, Murat Yüksek Lisans, Elektrik ve Elektronik Mühendisliği Bölümü Tez Yöneticisi: Assoc. Prof. Dr. Tolga Çiloğlu Ortak Tez Yöneticisi: Prof. Dr. Mete Severcan Eylül 2000, 55 sayfa Bu çalışma, gürültülü konuşmanın anlaşılabilirliğini arttıracak bir iyileştirme al goritmasının geliştirilmesini amaçlamaktadır. Hem renkli hem de beyaz gürültü etkisi için geçerli olan, parametrik sinyal modeline dayalı bir algoritma gelişti rilmiştir. Konuşmanın seslilik yapısını farklı frekans bantlarında öne çıkartan bir konuşma modeli oluşturulmuştur. Konuşmanın uyarım sinyali peryodik ve gürültülü sinyal karışımı olarak ele alınmış ve karışım oranları farklı bant aralıkları için ayrı ayrı hesaplanmıştır. Önerilen uyarım sinyal modeli ile konuşma için bir durum uzayı modeli oluşturulmuştur. Varsayılan model kullanılarak bir Kalman süzgeci tasarlanmış ve gürültülü gözlemlerden, temiz konuşma kestirilmeye ça-lışılmıştır. Sonuçlar, Kalman süzgeci ile iyileştirme yapan diğer algoritmalarla karşılaştırılmış, modellemenin ve kullanılan filtreleme algoritmasının etkileri tar tışılmıştır. Anahtar Kelimeler: Kalman süzgeci, konuşma iyileştirme vı ABSTRACT ENHANCEMENT OF NOISY SPEECH UTILIZING THE KALMAN FILTER Deviren, Murat M.S., Department of Electrical and Electronics Engineering Supervisor: Assoc. Prof. Dr. Tolga Çiloğlu Co-Supervisor: Prof. Dr. Mete Severcan September 2000, 55 pages The aim of this work is to develop an enhancement algorithm to increase the intelligibility of noisy speech. A parametric model based time domain algorithm is developed to filter out both white and colored noise interference. We devise a speech model that emphasizes the voicing characteristics in different frequency intervals. Excitation signal for speech is modelled as a mixture of periodic and white noise signals where the mixture proportions are calculated separately for different frequency intervals. Based on the proposed excitation model, we formu late a state space model for speech. The presumed speech model is then used to construct an optimal Kalman filter for estimation of clean speech from the noisy mobservations. The results are compared to other Kalman filter based speech en hancement algorithms. The effect of modelling and type of filtering is discussed. Keywords: Kalman filter, speech enhancement IV 55