A large Greek-English Dictionary with 81,515 entries, 192,592 translations into English and 50,106 usage examples with their translation has been developed in combined printed and electronic (DVD) form. The electronic dictionary features unique facilities for searching the entire or any part of the Greek and English section, and has incorporated a series of speech and language processing tools which efficiently assist learners of Greek and English. This paper presents the human-machine interface of the dictionary and the most important tools, i.e. the TTS-synthesizers for Greek and English, the lemmatizers for Greek and English, the Grapheme-to-Phoneme converter for Greek and the syllabification system for Greek.

Abstract Attempting to automatically learn to identify verb complements from natural language corpora without the help of sophisticated linguistic resources like grammars, parsers or treebanks leads to a significant amount of noise in the data. In machine learning terms, where learning from examples is performed using class-labelled feature-value vectors, noise leads to an imbalanced set of vectors: assuming that the class label takes two values (in this work complement/non-complement), one class (complements) is heavily underrepresented in the data in comparison to the other. To overcome the drop in accuracy when predicting instances of the rare class due to this disproportion, we balance the learning data by applying one-sided sampling to the training corpus and thus by reducing the number of non-complement instances. This approach has been used in the past in several domains (image processing, medicine, etc) but not in natural language processing. For identifying the examples that are safe to remove, we use the value difference metric, which proves to be more suitable for nominal attributes like the ones this work deals with, unlike the Euclidean distance, which has been used traditionally in one-sided sampling. We experiment with different learning algorithms which have been widely used and their performance is well known to the machine learning community: Bayesian learners, instance-based learners and decision trees. Additionally we present and test a variation of Bayesian belief networks, the COr-BBN (Class-oriented Bayesian belief network). The performance improves up to 22% after balancing the dataset, reaching 73.7% f-measure for the complement class, having made use only a phrase chunker and basic morphological information for preprocessing.

In this paper we cope with the task of modeling phoneme duration for Greek speech synthesis. In particular we apply well established machine learning approaches to the WCL-1 prosodic database for predicting segmental durations from shallow morphosyntactic and prosodic features. We employ decision trees, instance based learning and linear regression. Trained on a 5500 word database, both CART and linear regression models proved to be the most effective in terms for the task with a root mean square error off 0. 0252 and 0.0251 respectively.

The general problem addressed in this paper is that of separating the voices of active moving speakers in the presence of background noise and moderate reverberation level in the acoustic field using a single microphone array. We adapt the multisensor multitarget tracking theory to the context of microphone arrays in order to form receptive beams that lock on each moving speaker on an extended time basis and therefore, achieve voice separation. Our approach: 1) incorporates kinematical information of speakers' movement by using an interacting multiple model (IMM) estimator per speaker in order to constrain the evolution of direction of arrival (DOA) measurements, which characterize various motions of the speakers, and 2) can directly account for measurement origin uncertainty, i.e., which measurement comes from which speaker, by using the probabilistic-data-association technique in conjunction with the IMM estimator. The effectiveness of the approach is illustrated by an extensive simulation study on tracking the DOAs of two speakers with crossing trajectories and three static speakers having a conversation with partially overlapping speech and long pauses

This paper reports on the evaluation of automatic prosodic phrase break assignment. We utilize two tree-structured predictors, the commonly used CART and a C4.5, to predict break placement from sequences of easy to extract shallow textual features. We are experimenting with two 500-utterance sized prosodic corpora developed by two Greek universities that originate from different domains in order to focus on the differences in prediction between generic and limited domain datasets. The evaluation shows that while the limited dataset achieves better accuracy than the generic one in the CART case, this difference is lowered with the introduction of C4.5. Minor breaks proved to be the most difficult class to predict in CART case, while we achieved a 50 % of improvement with C4.5. 1.

One of the main aspects in Text-to-Speech (TtS) synthesis is the successful prediction of tonal events. In this work we deal with the evaluation of corpus-based models in operational environments other than the training ones. Two pitch accent frameworks derived by linguistically enriched speech data from a generic domain and a limited domain were initially evaluated by applying the 10-fold cross validation method. As a second step, we utilized the cross domains data validation. Due to the heterogeneity of the data, we further employed three machine learning approaches, CART, Naive Bayes and Bayesian networks. The results demonstrate that the limited domain models achieve in average 10 % improved accuracy in self-domain evaluation, while the generic models preserve a their performance regardless the domain of application. 1.

In this paper a handwritten recognition algorithm based on structural characteristics, histograms and profiles, is presented. The well-known horizontal and vertical histograms are used, in combination with the newly introduced radial histogram, out-in radial and in-out radial profiles for representing 32 /spl times/ 32 matrices of characters, as 280-dimension vectors. The recognition process has been supported by a lexical component based on dynamic acyclic FSAs (Finite-State-Automata).

We present an integrated system that uses speech as a natural input modality to provide user-friendly access to information and entertainment devices installed in a real home environment. The system is based on a combination of beamforming techniques and speech recognition. The general problem addressed in this work is that of hands-free speech recognition in a reverberant room where users walk while engaged in conversation in the presence of different types of house-specific noisy conditions (e.g. TV/radio broadcast, interfering speakers, ventilator/air-condition noise, etc). The paper focuses on implementation details and practical considerations concerning the integration of diverse technologies into a working system. 1.

For the present paper, a Bayesian probabilistic framework for the task of automatic acquisition of intonational phrase breaks was established. By considering two different conditional independence assumptions, the naïve Bayes and Bayesian networks approaches were regarded and evaluated against the CART algorithm, which has been previously used with success. A finite length window of minimal morphological and syntactic resources was incorporated, i.e. the POS label and the kind of phrase boundary, a novel syntactic feature that has not been applied to intonational phrase break detection before. This feature can be used in languages where syntactic parsers are not available and proves to be important, not only for the proposed Bayesian methodologies but for other algorithms, like CART. Trained on a 5500 word database, Bayesian networks proved to be the most effective in terms of precision (82,3%) and recall (77,2%) for predicting phrase breaks. 1.

In this paper we attempt to face common problems of handwritten documents such as nonparallel text lines in a page, hill and dale writing, slanted and connected characters. Towards this end an integrated system for document image preprocessing is presented. This system consists of the following modules: skew angle estimation and correction, line and word segmentation, slope and slant correction. The skew angle correction, slope correction and slant removing algorithms are based on a novel method that is a combination of the projection profile technique and the Wigner–Ville distribution. Furthermore, the skew angle correction algorithm can cope with pages whose text line skew angles vary, and handle them by areas. Our system can be used as a preprocessing stage to any handwriting character recognition or segmentation system as well as to any writer identification system. It was tested in a wide variety of handwritten document images of unconstrained English and Modern Greek text from about 100 writers. Additionally, combinations of the above algorithms have been used in the framework of the ACCeSS system (European project LE-1 1802, aiming at the automatic processing of application forms of insurance companies) as well as in the processing of GRUHD and IAM-B databases for automating the procedure of extracting data.

Our work introduces a speech enhancement technique that can explicitly incorporate prior information about the gender or speaker time-frequency characteristics in its formalism. We approximate the multimodal, clean speech linear spectrum magnitude with a mixture of Gaussians pdfs using the Expectation-Maximization algorithm (EM). Subsequently. we apply the Bayesian inference framework to the degraded spectral coefficients and by employing Minimum Mean Square Error Estimation (MMSE) we derive a closed fonn solution for the spectral magnitude estimation task adapted to the spectral characteristics and noise variance of each band. We suggest that 2–3 minutes of phonetically balanced non-degraded gender or speaker dependent speech is adequate to tune our algorithm. We demonstrate the benefit of using an enhancement technique tailored to a specific gender or speaker and propose its use in cases where message ambiguity is of critical importance. We evaluate of our algorithm using Lynx helicopter and White Gaussian noise on the task of improving the quality of speech and in combination with a speech coder and demonstrate its robustness at very low SNRs. Implementation code is available at: http://slt.wcl.ee.upatras.gr/potamitis/index.html

This paper presents and analyzes an incremental algorithm for the construction of Acyclic Non-deterministic Finite-state Automata (NFA). Automata of this type are quite useful in computational linguistics, especially for storing lexicons. The proposed algorithm produces compact NFAs, i.e. NFAs that do not contain equivalent states. Unlike Deterministic Finite-state Automata (DFA), this property is not sufficient to ensure minimality, but still the resulting NFAs are considerably smaller than the minimal DFAs for the same languages.

Our work introduces the sparse code shrinkage (SCS) technique as a speech enhancement algorithm that aims at improving the quality of speech perception. SCS is a fairly new statistical technique originally presented to the applied mathematics and image denoising community, but, to our knowledge, its potential for speech enhancement has not yet been exploited. Its application on speech denoising gives rise to a conceptual framework which is quite different from the techniques dominating the speech enhancement domain. SCS originates in applying independent component analysis (ICA) to a large ensemble of clean speech frames, revealing their underlying basis of statistically independent functions. Projecting the frames composing a noisy speech signal on this basis, facilitates the application of Bayesian denoising to each of the resulting independent components individually. The maximum a posteriori (MAP) formulation leads to a soft threshold function optimally adapted to the statistics of each independent component which effectively reduces white and coloured Gaussian noise. Subsequently, an inverse transformation from the ICA-transformed domain back to the time domain reconstructs the enhanced signal.

We present the GRUHD database of Greek characters, text, digits and other symbols in unconstrained handwriting mode. The database consists of 1,760 forms that contain 667,583 handwritten symbols and 102,692 words in total, written by 1,000 writers, 500 men and equal number of women. Special attention was paid to gathering data from writers of different age and educational level. The GRUHD database is accompanied by the GRUHD software that facilitates its installation and use and enables the user to extract and process the data from the forms selectively, depending on the application. The various types of possible installations make it appropriate for the training and validation of character recognition, character segmentation and text-dependent writer identification systems.

This paper deals with the use of Bayesian Belief Networks in order to improve the accuracy and training time of character segmentation for unconstrained handwritten text. Comparative experimental results have been evaluated against Naive Bayes classification, which is based on the assumption of the independence of the parameters and two additional previous commonly used methods. Results have depicted that obtaining the inferential dependencies of the training data, could lead to the reduction of the required training time and size by a factor of 55%. Moreover, the achieved accuracy in detecting segment boundaries exceeds 86% whereas limited training data are proved to endow with very satisfactory results.

Learning Bayesian belief networks (BBN) from corpora and support vector machines (SVM) have been applied to the automatic acquisition of verb subcategorization frames for Modern Greek. We are incorporating minimal linguistic resources, i.e. basic morphological tagging and phrase chunking, to demonstrate that verb subcategorization, which is of great significance for developing robust natural language human computer interaction systems, could be achieved using large corpora, without having any general-purpose, syntactic parser at all. In addition, apart from BBN and SVM, which have not previously used for this task, we have experimented with three well-known machine learning methods (feedforward backpropagation neural networks, learning vector quantization and decision tables), which are also being applied to the task of verb subcategorization frame defection for the first time. We argue that both BBN and SVM are well suited for learning to identify verb subcategorization frames. Empirical results will support this claim. Performance has been methodically evaluated using two different corpora types, one balanced and one domain-specific in order to determine the unbiased behaviour of the trained models. Limited training data are proved to endow with satisfactory results. We have been able to achieve precision exceeding 80% on the identification of subcategorization frames which were not known beforehand.

A novel method for accurate detection of regions of interest (ROIs) that contain circumscribed lesions in mammograms is presented. The mammograms are segmented using a statistical threshold and a number of candidate regions are extracted. Then a set of qualification criteria is employed to filter these regions retaining the most suspicious for which a radial-basis function neural network makes the final decision marking them as ROIs that contain abnormal tissue. The proposed method detects the exact location of the circumscribed lesions with an accuracy of 90.9%, and a very low number of false positive regions per image (2.1 ROIs per image) in the MIAS database.

The classical method of document skew detection, based on nearest-neighbor clustering, is revisited. A heuristic is proposed which attempts to group all the connected components that belong to the same line of text, into one cluster. The larger clusters are known to result in better skew angle estimation. The skew detection accuracy of this improved connected-components method is several orders of magnitude better, when compared to the classical approach, with no change in the order of complexity.

In this paper we present a novel solution to the Overdetermined Blind Speech Separation (OBSS) problem for improving speech recognition accuracy of N simultaneous speakers in real room environments using M, (M>N) microphones. The proposed OBSS system uses basic NxN Blind Speech Separation networks that process in parallel all different combinations of the available mixture signals in the frequency domain, resulting to lower computational complexity and faster convergence. Extensive experiments using an array of two to ten microphones and two simultaneous speakers in a simulated real room, showed that when the number of the microphones increases beyond two, the separation performance is improved and the phoneme recognition accuracy of an HMM based decoder increases drastically (more than 6%). Therefore, the introduction of more microphones than speakers is justified in order to improve speech recognition accuracy in multi simultaneous speaker environments.

In this paper we present a new on-line blind signal separation method capable to separate convolutive speech signals of moving speakers in highly reverberant rooms. The separation network used is a recurrent network which performs separation of convolutive speech mixtures in the time domain, without any prior knowledge of the propagation media, based on the maximum likelihood estimation (MLE) principle. The proposed method proved to be able to improve significantly (more than 10% in all adverse mixing situations) the performance of a continuous phoneme-based speech recognition system and therefore can be used as a front-end to separate simultaneous speech of speakers who are moving in arbitrary directions in reverberant rooms.

This article presents a novel technique for suppressing the effect of impulsive noise in the context of automatic speech recognition (ASR). The noise suppression scheme is based on the cooperation of two neural networks. The first one is responsible for the detection of corrupted cepstral vectors due to the presence of impulsive noise. The second one is dedicated to the restoration of the detected problematic vectors solely. The novelty of the method lies in the robust detection of impulsive noise regardless of the noise source and the local restoration of the feature vectors. By avoiding a global act on the original waveform as spectral subtraction and Wiener filters do, we don't inflict any distortions on an already clean part of the original waveform. Extensive experimental valuation on a spoken digit database in the presence of machine gun impulsive noise has proved the robustness of our method.

Presents a character segmentation algorithm for unconstrained cursive handwritten text. The transformation-based learning method and a simplified variation of it are used in order to extract automatically rules that detect the segment boundaries. Comparative experimental results are given for a collection of multiwriter handwritten words. The achieved accuracy in detecting segment boundaries exceeds 82%. Moreover limited training data can provide very satisfactory results.

We describe a novel technique for improving speech recognition performance in real environments. We investigate the special case of speech recognition in the car environment for SNRs ranging from -10 to 20 dB. Our approach makes use of a feature set that is composed of uncorrelated variables in order to create a group of neural networks each one dedicated to a sole variable of the feature vector. This technique results in neural networks of much smaller total number of weights than reported cases and consequently in faster training and execution performance. Furthermore, contextual information regarding a feature's history is incorporated into the network by making use of recurrent neural networks. We evaluate the performance in comparison with the standard MLPs and TDNNs in order to prove that they compare favourably to them in terms of recognition improvement over a wide range of SNRs.

The present paper addresses the question of the efficiency of Independent Component Analysis (ICA) as a statistical process for deriving optimal representational bases for the projection of spectrum and cepstrum in the context of Automatic Speech Recognition (ASR). Several decorrelation strategies have been applied on the log-spectrum and cepstrum to fulfill the practical need of a diagonal covariance HMM for uncorrelated features. In our work we question the optimality of a fixed decorrelation strategy as DCT and follow an emerging trend in ASR that designs projection bases based on the statistics of speech. We differentiate our approach from the second order statistics of

The two main factors that characterize a text are its content and its style, and both can be used as a means of categorization. In this paper we present an approach to text categorization in terms of genre and author for Modern Greek. In contrast to previous stylometric approaches, we attempt to take full advantage of existing natural language processing (NLP) tools. To this end, we propose a set of style markers including analysis-level measures that represent the way in which the input text has been analyzed and capture useful stylistic information without additional cost. We present a set of small-scale but reasonable experiments in text genre detection, author identification, and author verification tasks and show that the proposed method performs better than the most popular distributional lexical measures, i.e., functions of vocabulary richness and frequencies of occurrence of the most frequent words. All the presented experiments are based on unrestricted text downloaded from the World Wide Web without any manual text preprocessing or text sampling. Various performance issues regarding the training set size and the significance of the proposed style markers are discussed. Our system can be used in any application that requires fast and easily adaptable text categorization in terms of stylistically homogeneous categories. Moreover, the procedure of defining analysis-level markers can be followed in order to extract useful stylistic information using existing text processing tools.

A novel method for fast detection of regions of suspicion (ROS) that contain circumscribed lesions in mammograms is presented. The position and the size of ROS are first recognized with the aid of a radial-basis-function neural network (RBFNN) by performing windowing analysis. Then a set of criteria is employed to these regions to make the final decision concerning the abnormal ones. Accelerated estimation of the high-order statistical features decreases the computational complexity 55 times in multiplication operations. The proposed method detects the exact location of the circumscribed lesions with accuracy of 72.7% (overlap between groundtruthed and detected regions greater than 50%) for mammograms containing masses, while the recognition rate for the normal ones reaches 77.7% in the MIAS database.

The aim of this paper is to study the use of a robust acquisition system based on a microphone array for speech related applications in real situations. A comparison is performed between two beamforming methods: the Delay and Sum beamforming (DS) and the Spatial Reference Optimal beamforming (SRO). Both of them are frequency domain designed, using harmonic spatial distributed microphones. The quality of the microphone array output signal for the different beamforming approaches is measured by using a continuous speech recognition system based on discrete HMMs. Results of the speech recognition system show the advantage of using a microphone array with the SRO beamforming as front-end.

This paper presents a hybrid ANN/HMM syllable recognition module based on vowel spotting. An advanced multi-level vowel spotting method is used to achieve minimum vowel loss and accurate detection of the vowel location and duration. Discrete Hidden Markov Models (DSHMM), Multi Layer Perceptrons (MLP) and Heuristics (HR) are used for this purpose. A hybrid ANN/HMM technique is then used to recognize the syllables between the detected vowels. We replace the usual DSHMM probability parameters with combined neural network outputs. For this purpose both context dependent (CD) and context independent (CI) neural networks are used. Global normalization is employed on the parameters as opposed to the local normalization used on parameters in standard HMMs. Also, all parameters are estimated simultaneously according to the discriminative conditional maximum likelihood (CML) criterion. The tests were performed on the TIMIT and NTIMIT databases and showed significant performance improvement compared to similar systems.

This paper describes a user-assisted business letter generator that meets the ever-increasing demand for more flexible and modular letter generators which draw on explicit thematic models and which are easily adaptable to specific user needs. Based on a detailed analysis of requirements and taking full advantage of end-user feedback, the presented generator not only creates a business letter according to the user's choices, it also refines it, taking into consideration stylistic aspects like written style and tone.

In extensive experiments, the recognition score of a speaker independent isolated word speech recognition system based on a continuous density HMM (CDHMM) has been measured in the presence of real life noises in various SNRs. In all experiments the results show improvement in the mean recognition score when the subband adaptive filtering LMS method is used in comparison to the full-band LMS method. This improvement increases when changing types of noise distort the speech signal.

A novel speech enhancement technique is presented based on the definition of the psychoacoustically derived quantity of audible noise spectrum and its subsequent suppression using optimal nonlinear filtering of the short-time spectral amplitude (STSA) envelope. The filter operates with sparse spectral estimates obtained from the STSA, and, when these parameters are accurately known, significant intelligibility gains, up to 40%, result in the processed speech signal. These parameters can be also estimated from noisy data, resulting into smaller but significant intelligibility gains.

The presented work is strongly motivated by the need for categorizing unrestricted text in terms of a functional style (FS) in order to attain a satisfying outcome in style processing. Towards this aim, a three level description of FS is given that comprises: (a) the basic categories of FS; (b) the main features that characterize each one of the above categories; and (c) the linguistic identifiers that act as style markers in text for the identification of the above features. Special emphasis is put on the problems that faced the computational implementation of the aforementioned findings, as well as the selection of the most appropriate stylometrics (i.e., stylistic scores) to achieve better results on text categorization. This approach is language independent, empirically driven, and can be used in various applications including grammar and style checking, natural language generation, style verification in real world text, and recognition of style shift between adjacent portions of text.

This paper presents a model for large-scale hierarchical networks which leads to an almost optimal path bandwidth (PB) allocation, under constraints posed by the installed bandwidth in the transmission links of the network. The PB allocation procedure consists of three phases. In the first phase, a suitable decomposition of the whole network takes place, in which the hierarchical network is split to a set of sub-networks of just one level. In the second phase, the optimization algorithm developed for one-level telecommunication networks is applied to each subnetwork to define the optimal PB allocation. The optimization criterion is to minimize the worst call blocking probability (CBP) of all switching pairs of the sub-network. In the third phase, composition of the sub-networks takes place in a successive way which leads to the final PB allocation. As the hierarchical network is built up from optimized sub-networks, it is considered that it has an almost optimal PB allocation. The proposed PB allocation scheme is explained by means of an example.

Operating system command languages assist the user in executing commands for a significant number of common everyday tasks. On the other hand, the introduction of textual command languages for robots has provided the opportunity to perform some important functions that leadthrough programming cannot readily accomplish. However, such command languages assume the user to be expert enough to carry out a specific task in these application domains. On the contrary, a natural language interface to such command languages, apart from being able to be integrated into a future speech interface, can facilitate and broaden the use of these command languages to a larger audience. In this paper, advanced techniques are presented for an adaptive natural language interface that can (a) be portable to a large range of command languages, (b) handle even complex commands thanks to an embedded linguistic parser, and (c) be expandable and customizable by providing the casual user with the opportunity to specify some types of new words as well as the system developer with the ability to introduce new tasks in these application domains. Finally, to demonstrate the above techniques in practice, an example of their application to a Greek natural language interface to the MS-DOS operating system is given.

This paper presents a clustering algorithm producing multiple whole-word continuous density hidden Markov models (CDHMM) for isolated word recognition systems. The algorithm estimates a minimum number of CDHMM per word that approaches or satisfies a minimum predefined word-dependent recognition accuracy in the training set. Significantly lower memory requirements and a better and more uniformly distributed recognition accuracy among the words of the vocabulary are measured by comparing this algorithm with the modified K-means clustering method.

Free word-order languages cause problems when being analysed due to their great composing power and flexibility. The paper deals with the problems stemming from the analysis of complex sentences as well as with the adoption of suitable grammars to parse them in free word-order languages. Based on a careful study of large Greek corpora, we propose a method for parsing complex sentences in such languages that combines heuristic information and pattern matching techniques in early processing levels. Although this method has been implemented and tested in the Greek language, its theoretical background and results are language independent and can be proved useful for many practical applications.>

The authors address the problem of script representation and script selection in the presence of weak semantic ambiguity and proposes some useful techniques to increase flexibility in current script-like structures. To highlight the proposed approaches, examples are given and the general outline of a prototype model, implemented in Prolog, is presented.

The authors present a simple and fast algorithm for accurately locating the endpoints of isolated words spoken in an office environment. This algorithm is based on energy measures and threshold logic and is adaptive to almost any low-noise acoustic environment. Experimental results from the algorithm's application were both acoustically checked and compared to hand-edited results by skilled personnel. The algorithm was tested on a vocabulary of 34 words, uttered by 25 speakers and recorded in different acoustic environments. The accuracy of the algorithm is quite acceptable.>

A modular software TTS (text-to-speech) system for Greek with good intelligibility and quality of speech and the possibility of being further improved by extending its linguistic knowledge is presented. The system has several peculiarities in comparison to most systems for other languages, combining the advantages of formant synthesis with those of diphone synthesis. In addition to the text normalizer (including numbers) and the sophisticated text preprocessor, the system uses composite speech segments besides phonemes which are concatenated, using a dynamically-adjusted-in-range, sigmoid function. The segments are coded in a novel scheme aiding the rules which manipulate voice onset and duration times. A declined line with its ending part dependent on the punctuation mark or function word, which fluctuates according to the stressed points and unvoiced (voiced) consonant and plosive locations, controls the intonation of the input text. The system is to be improved by incorporating elaborate prosodic rules resulting from syntactic analysis of the text.>