Pattern recognition tasks often face the situation that training data are not fully representative of test data. This problem is well-recognized in speech recognition, where methods like cepstral mean normalization (CMN), vocal tract length normalization (VTLN) and maximum likelihood linear regression (MLLR) are used to compensate for channel and speaker differences. Speech emotion recognition (SER) is an important emerging field in human-computer interaction and faces the same data shift problems, a fact which has been generally overlooked in this domain. In this paper, we show that compensating for channel and speaker differences can give significant improvements in SER by modelling these differences as a covariate shift. We employ three algorithms from the domain of transfer learning that apply importance weights (IWs) within a support vector machine classifier to reduce the effects of covariate shift. We test these methods on the FAU Aibo Emotion Corpus, which was used in the Interspeech 2009 Emotion Challenge. It consists of two separate parts recorded independently at different schools; hence the two parts exhibit covariate shift. Results show that the IW methods outperform combined CMN and VTLN and significantly improve on the baseline performance of the Challenge. The best of the three methods also improves significantly on the winning contribution to the Challenge.

This paper extends binary support vector machines to multiclass classification for recognising emotions from speech.We apply two standard schemes (one-versus-one and one-versusrest) and two schemes that form a hierarchy of classifiers each making a distinct binary decision about class membership, on three publicly-available databases.Using the OpenEAR toolkit to extract more than 6000 features per speech sample, we have been able to outperform the state-of-the-art classification methods on all three databases.

The mean of a data set is one trivial representation of data from one class. Recently, mutual interdependence analysis (MIA) has been successfully used to extract more involved representations, or ldquomutual featuresrdquo, accounting for samples in the class. For example a mutual feature is a speaker signature under varying channel conditions or a face signature under varying illumination conditions. A mutual representation is a linear regression that is equally correlated with all samples of the input class. We present the MIA optimization criterion from the perspectives of regression, canonical correlation analysis and Bayesian estimation. This allows us to state and solve the above criterion concisely, to contrast the MIA solution to the sample mean, and to infer other properties of its closed form, unique solution under various statistical assumptions. We define a generalized MIA solution (GMIA) and apply MIA and GMIA in a text-independent speaker verification task using the NTIMIT database. Both methods show competitive performance with equal-error-rates of 7.5 % and 6.5 % respectively over 630 speakers.

We describe extensive experiments on the recognition of emotion from speech using acoustic features only. Two databases of acted emotional speech (Berlin and DES) have been used in this work. The principal focus is on methods for selection of good features from a relatively large set of hand-crafted features, perhaps formed by fusing different feature sets used by different researchers. We show that the monotonic assumption underlying popular sequential selection algorithms does not hold, and use this finding to improve recognition accuracy. We show further that a very simple classifier (k-nearest neighbour) produces better results than any so far reported by other researchers on these databases, suggesting that previous work has failed to match the complexity of the classifier used to the complexity of the data. Finally, several potentially fruitful avenues for future work are outlined.

One does not have to go very far into the subject of artificial intelligence (AI) before deep philosophical issues start to surface. The Turing test, the frame problem, the Chinese room argument, symbol grounding, the role of representation: these are just a few of the topics that have generated intense discussion and debate over the years. One particular controversy surrounds symbolism: whether or not it is necessary to have explicit symbol representation and, if so, the mechanism(s) by which these symbols come into existence. How are such questions to be answered? Generally, they are considered to lie in the realm of philosophy and not to be amenable to experimental tests. However, when even limited experimental testing is feasible, it can offer valuable insight. In this article, we present an empirical test designed to explore some important foundational issues in AI. An artificial agent inhabits a digital world (a cellular automaton) in which its cognitive abilities vary in three dimensions (size of symbol memory, percentage of symbols that are innate, planning depth), allowing us to position it in a space that reflects degree of commitment to key philosophical standpoints. One plane of this space corresponds to pure symbol attachment, another plane corresponds to pure symbol grounding, and the origin of coordinates corresponds to pure enactivism. We find that an enactivist (purely reactive) agent architecture can perform as well as one employing planning in this scenario, if properly designed. Planning has strengths when task/environment complexity make design difficult but weaknesses if an inappropriate world model is acquired (e.g. as a result of mismatch between model and task/environment complexity). However, the main claim of the article is that empirical exploration of the kind presented here could usefully form the initial phase of the design of many practical AI systems, and forms a valuable alternative to simply declaring a priori adherence to a particular philosophical position.

Noisy or distorted video/audio training sets represent constant challenges in automated identification and verification tasks. We propose the method of Mutual Interdependence Analysis (MIA) to extract "mutual features" from a high dimensional training set. Mutual features represent a class of objects through a unique direction in the span of the inputs that minimizes the scatter of the projected samples of the class. They capture invariant properties of the object class and can therefore be used for classification. The effectiveness of our approach is tested on real data from face and speaker recognition problems. We show that "mutual faces" extracted from the Yale database are illumination invariant, and obtain identification error rates of 2.2% in leave-one-out tests for differently illuminated images. Also, "mutual speaker signatures" for text independent speaker verification achieve state-of-the- art equal error rates of 6.8% on the NTIMIT database.

Abstract Automatic pronunciation of unknown words (i.e., those not in the system dictionary) is a difficult problem in text-to-speech (TTS) synthesis. Currently, many data-driven approaches have been applied to the problem, as a backup strategy for those cases where dictionary matching fails. The difficulty of the problem depends on the complexity of spelling-to-sound mappings according to the particular writing system of the language. Hence, the degree of success achieved varies widely across languages but also across dictionaries, even for the same language with the same method. Further, the sizes of the training and test sets are an important consideration in data-driven approaches. In this paper, we study the variation of letter-to-phoneme transcription accuracy across seven European languages with twelve different lexicons. We also study the relationship between the size of dictionary and the accuracy obtained. The largest dictionaries of each language have been partitioned into ten approximately equal-sized subsets and combined to give ten different-sized test sets. In view of its superior performance in previous work, the transcription method used is pronunciation by analogy (PbA). Best results are obtained for Spanish, generally believed to have a very regular (‘shallow’) orthography, and poorest results for English, a language whose irregular spelling system is legendary. For those languages for which multiple dictionaries were available (i.e., French and English), results were found to vary across dictionaries. For the relationship between dictionary size and transcription accuracy, we find that as dictionary size grows, so performance grows monotonically. However, the performance gain decelerates (tends to saturate) as the dictionary increases in size; the relation can simply be described by a logarithmic regression, one parameter of which (α) can be taken as quantifying the depth of orthography of a language. We find that α for a language is significantly correlated with transcription performance on a small dictionary (approximately 10,000 words) for that language, but less so for asymptotic performance. This may be because our measure of asymptotic performance is unreliable, being extrapolated from the fitted logarithmic regression.

Previous research applying kernel methods such as support vector machines (SVMs) to hyperspectral image classification has achieved performance competitive with the best available algorithms. However, few efforts have been made to extend SVMs to cover the specific requirements of hyperspectral image classification, for example, by building tailor-made kernels. Observation of real-life spectral imagery from the AVIRIS hyperspectral sensor shows that the useful information for classification is not equally distributed across bands, which provides potential to enhance the SVM's performance through exploring different kernel functions. Spectrally weighted kernels are, therefore, proposed, and a set of particular weights is chosen by either optimizing an estimate of generalization error or evaluating each band's utility level. To assess the effectiveness of the proposed method, experiments are carried out on the publicly available 92AV3C dataset collected from the 220-dimensional AVIRIS hyperspectral sensor. Results indicate that the method is generally effective in improving performance: spectral weighting based on learning weights by gradient descent is found to be slightly better than an alternative method based on estimating "relevance" between band information and ground truth.

In spite of difficulty in defining the syllable unequivocally, and controversy over its role in theories of spoken and written language processing, the syllable is a potentially useful unit in several practical tasks which arise in computational linguistics and speech technology. For instance, syllable structure might embody valuable information for building word models in automatic speech recognition, and concatenative speech synthesis might use syllables or demisyllables as basic units. In this paper, we first present an algorithm for determining syllable boundaries in the orthographic form of unknown words that works by analogical reasoning from a database or corpus of known syllabifications. We call this syllabification by analogy (SbA). It is similarly motivated to our existing pronunciation by analogy (PbA) which predicts pronunciations for unknown words (specified by their spellings) by inference from a dictionary of known word spellings and corresponding pronunciations. We show that including perfect (according to the corpus) syllable boundary information in the orthographic input can dramatically improve the performance of pronunciation by analogy of English words, but such information would not be available to a practical system. So we next investigate combining automatically-inferred syllabification and pronunciation in two different ways: the series model in which syllabification is followed sequentially by pronunciation generation; and the parallel model in which syllabification and pronunciation are simultaneously inferred. Unfortunately, neither improves performance over PbA without syllabification. Possible reasons for this failure are explored via an analysis of syllabification and pronunciation errors.

Automatic pronunciation of unknown words of English is a hard problem of great importance in speech technology. Proper names constitute an especially difficult class of words to pronounce because of their variable origin and uncertain degree of assimilation of foreign names to the conventions of the local speech community. In this paper, we compare four different methods of proper name pronunciation for English text-to-speech (TTS) synthesis. The first (intended to be used as the primary strategy in a practical TTS system) uses a set of manually supplied pronunciations, referred to as the “dictionary” pronunciations. The remainder are pronunciations obtained from three different data-driven approaches (intended as candidates for the back-up strategy in a real system) which use the dictionary of “known” proper names to infer pronunciations for unknown names. These are: pronunciation by analogy (PbA), a decision tree method (CART), and a table look-up method (TLU). To assess the acceptability of the pronunciations to potential users of a TTS system, subjective evaluation was carried out, in which 24 listeners rated 1200 synthesized pronunciations of 600 names by the four methods using a five-point (opinion score) scale. From over 50 000 proper names and their pronunciations, 150 so-calledone-of-a-kindpronunciations were selected for each of the four methods (600 in total). Aone-of-a-kindpronunciation is one for which one of the four methods disagrees with the other three methods, which agree among themselves. Listener opinions onone-of-a-kindpronunciations are argued to be a good measure of the overall quality of a particular method. For eachone-of-a-kindpronunciation, there is a corresponding so-calledrestpronunciation (another 600 in total), on which the remaining three competitor methods agree, for which listener opinions are taken to be indicative of the general quality of the competition. Nonparametric tests of significance of mean opinion scores show that the dictionary pronunciations are rated superior to the automatically inferred pronunciations with little difference between the data-driven methods for theone-of-a-kindpronunciations, but for therestpronunciations there is suggestive evidence that PbA is superior to both CART and TLU, which perform at approximately the same level.

Spectral band selection is a fundamental problem in hyperspectral data processing. In this letter, a new band-selection method based on mutual information (MI) is proposed. MI measures the statistical dependence between two random variables and can therefore be used to evaluate the relative utility of each band to classification. A new strategy is described to estimate the MI using a priori knowledge of the scene, reducing reliance on a "ground truth" reference map, by retaining bands with high associated MI values (subject to the so-called "complementary" conditions). Simulations of classification performance on 16 classes of vegetation from the AVIRIS 92AV3C data set show the effectiveness of the method, which outperforms an MI-based method using the associated reference map, an entropy-based method, and a correlation-based method. It is also competitive with the steepest ascent algorithm at much lower computational cost

Extracting full-body motion of walking people from monocular video sequences in complex, real-world environments is an important and difficult problem, going beyond simple tracking, whose satisfactory solution demands an appropriate balance between use of prior knowledge and learning from data. We propose a consistent Bayesian framework for introducing strong prior knowledge into a system for extracting human gait. In this work, the strong prior is built from a simple articulated model having both time-invariant (static) and time-variant (dynamic) parameters. The model is easily modified to cater to situations such as walkers wearing clothing that obscures the limbs. The statistics of the parameters are learned from high-quality (indoor laboratory) data and the Bayesian framework then allows us to "bootstrap" to accurate gait extraction on the noisy images typical of cluttered, outdoor scenes. To achieve automatic fitting, we use a hidden Markov model to detect the phases of images in a walking cycle. We demonstrate our approach on silhouettes extracted from fronto-parallel ("sideways on") sequences of walkers under both high-quality indoor and noisy outdoor conditions. As well as high-quality data with synthetic noise and occlusions added, we also test walkers with rucksacks, skirts, and trench coats. Results are quantified in terms of chamfer distance and average pixel error between automatically extracted body points and corresponding hand-labeled points. No one part of the system is novel in itself, but the overall framework makes it feasible to extract gait from very much poorer quality image sequences than hitherto. This is confirmed by comparing person identification by gait using our method and a well-established baseline recognition algorithm.

Automatic pronunciation of unknown words is a hard problem of great importance in speech technology. Proper names constitute an especially difficult class of words to pronounce because of their low frequency of occurrence and variable origin. In this paper, we compare three different data-driven approaches which use a dictionary of (known) proper names to infer pronunciations for unknown names, namely: pronunciation by analogy (PbA), the table look-up method by Weijters, and the ‘improved’ table look-up method by Daelemans and van den Bosch. Evaluation is both objective, in which inferred pronunciations are compared with ‘gold standard’ dictionary pronunciations, and subjective, in which listeners rated synthesised pronunciations using a 5-point scale. Objective evaluation used a leave-one-out technique on 52,911 names in the CMUDICT dictionary. Results show that PbA achieves the best performance at 63.93% names correct. In the subjective evaluation of 529 different pronunciations of 200 names, 12 listeners rated the pronunciations. Non-parametric tests of significance show that the dictionary pronunciations are rated superior to the automatically-inferred pronunciations; PbA is superior to both table look-up methods, and the ‘improved’ table look-up is superior to Weijters’ original method (Walsh test, p < 0.005 in all cases.)

Abstract The extensible markup language XML can be used to support the integration of several component programming environments to create a flexible physical simulation system. Data exchange via open‐standard‐based plain text files allows system components to be loosely‐coupled, rather than combined into an integrated development environment, so that the most appropriate tools can be used for each component and the system can be extended with minimal disruption. This paper details an example application using this technology to configure a simulation of robotic manipulation. Those parts of the system that require real‐time data exchange use simple UNIX socket‐based interactions, which are configured using shared XML setup files. The approach provides a reusable template for other, similar projects. Copyright © 2004 John Wiley & Sons, Ltd.

This paper describes a robust algorithm for arbitrary object tracking in long image sequences. This technique extends the dynamic Hough transform proposed in our earlier work to detect arbitrary shapes undergoing affine motion. The proposed tracking algorithm processes the whole image sequence globally. First, the object boundary is represented in lookup-table form, and we then perform an operation that estimates the energy of the motion trajectory in the parameter space. We assign an extra term in our cost function to incorporate smoothness of deformation. The object is actually rigid, so by 'deformation' we mean changes due to rotation or scaling of the object. There is no need for training or initialization, and an efficient implementation can be achieved with coarse-to-fine dynamic programming and pruning. The method, because of its evidence-based nature, is robust under noise and occlusion.

This paper proposes active handwriting models, in which kernel principal component analysis is applied to capture nonlinear handwriting variations. In the recognition phase, the chamfer distance transform and a dynamic tunneling algorithm (DTA) are employed to search for the optimal shape parameters. The proposed methodology is successfully applied to a novel radical decomposition approach to the challenging problem of handwritten Chinese character recognition.

This paper described a robust algorithm for arbitrary object tracking in long image sequences. This technique extends the dynamic Hough transform proposed in our earlier work to detect arbitrary shapes undergoing affine motion. The proposed tracking algorithm processes the whole image sequence globally. First, the object boundary is represented in lookup-table form, and we then perform an operation that estimates the energy of the motion trajectory in the parameter space. We assign an extra term in our cost function to incorporate smoothness of deformation. The object is actually rigid, so by 'deformation' we mean changes due to rotation or scaling of the object. There is no need for training or initialization, and an efficient implementation can be achieved with coarse-to-fine dynamic programming and pruning. The method, because of its evidence-based nature, is robust under noise and occlusion.

Handwritten Chinese characters can be recognized by first extracting the basic shapes (radicals) of which they are composed. Radicals are described by nonlinear active shape models and optimal parameters found using the chamfer distance transform and a dynamic tunneling algorithm. The radical recognition rate is 96.5 percent correct (writer-independent) on 280,000 characters containing 98 radical classes.

Offline handwritten Chinese character recognition is a very hard pattern-recognition problem of considerable practical importance. Two popular approaches are to extract features holistically from the character image or to decompose characters structurally into component parts---usually strokes. Here we take a novel approach, that of decomposing into radicals on the basis of image information (i.e., without first decomposing into strokes). During training, 60 examples of each radical were represented by "landmark" points, labeled semiautomatically, with radicals in different characteristic positions treated as distinctly different radicals. Kernel principal-component analysis then captured the main (nonlinear) variations around the mean radical. During the recognition, the dynamic tunneling algorithm was used to search for optimal shape parameters in terms of chamfer distance minimization. Considering character composition as a Markov process in which up to four radicals are combined in some assumed sequential order, we can recognize complete, hierarchically-composed characters by using the Viterbi algorithm. This gave a character recognition rate of 93.5% characters correct (writer-independent) on a test set of 430,800 characters from 2,154 character classes composed of 200 radical categories, which is comparable to the best reported results in the literature. Although the initial semiautomatic landmark labeling is time consuming, the decomposition approach is theoretically well-motivated and allows the different sources of variability in Chinese handwriting to be handled separately and by the most appropriate means--either learned from example data or incorporated as prior knowledge. Hence, high generalizability is obtained from small amounts of training data, and only simple prior knowledge needs to be incorporated, thus promising robust recognition performance. As such, there is very considerable potential for further development and improvement in the direction of larger character sets and less constrained writing conditions.

The traveling salesman problem (TSP) is a prototypical problem of combinatorial optimization and, as such, it has received considerable attention from neural-network researchers seeking quick, heuristic solutions. An early stage in many computer vision tasks is the extraction of object shape from an image consisting of noisy candidate edge points. Since the desired shape will often be a closed contour, this problem can be viewed as a version of the TSP in which we wish to link only a subset of the points/cities (i.e. the "noisefree" ones). None of the extant neural techniques for solving the TSP can deal directly with this case. In this paper, we present a simple but effective modification to the (analog) elastic net of Durbin and Willshaw which shifts emphasis from global to local behavior during convergence, so allowing the net to ignore some image points. Unlike the original elastic net, this semi-localized version is shown to tolerate considerable amounts of noise. As an example practical application, we describe the extraction of "pseudo-3D" human lung outlines from multiple preprocessed magnetic resonance images of the torso. An effectiveness measure (ideally zero) quantifies the difference between the extracted shape and some idealized shape exemplar. Our method produces average effectiveness scores of 0.06 for lung shapes extracted from initial semi-automatic segmentations which define the noisefree case. This deteriorates to 0.1 when extraction is from a noisy edge-point image obtained fully-automatically using a feedforward neural network.

This paper applies active handwriting models (AHM) to handwritten Chinese character recognition. Exploiting active shape models (ASM), the AHM can capture the handwriting variation from character skeletons. The AHM has the following characteristics: principal component analysis is applied to capture variations caused by handwriting, an energy functional on the basis of chamfer distance transform is introduced as a criterion to fit the model to a target character skeleton, and the dynamic tunneling algorithm (DTA) is incorporated with gradient descent to search for shape parameters. The AHM is used within a radical approach to handwritten Chinese characters recognition, which converts the complex pattern recognition problem to recognizing a small set of primitive structures-radicals. Our initial experiments are conducted on 98 radicals covering 1400 loosely-constrained Chinese character categories written by 200 different writers. The correct matching rate is 94.2% on these 2.8/spl times/10/sup 5/ characters. Comparison with existing radical approaches shows that our method achieves superior performance.

Handwritten Chinese character recognition is one of the most difficult problems of pattern recognition. Since the majority of Chinese characters are made up from just a small set of primitive structures (radicals), this paper describes an approach to active radical modeling for such handwritten characters. The most significant characteristic of our method is that radicals can be found robustly without stroke extraction, and the principal variations of the radical can be encoded in a small number of parameters. In the training phase, the example radicals are represented by manually-labeled 'landmark' points. Then a small number of principal components of the eigenvectors are calculated to capture the main variation of the training examples from the mean radical. In the matching phase, each radical model is fitted to the image evidence by adjusting the shape parameters in terms of chamfer distance minimization. Initial experiments are conducted on 1,100 loosely-constrained Chinese character categories written by 200 different writers. The correct matching rate is 95.8%, showing that our radical modeling is effective and capable of forming a sound basis for handwritten Chinese character recognition.

Motion tracking is an important task in computer vision. A new technique, the dynamic velocity Hough transform (DVHT), for tracking of parametric objects is described that extends the velocity Hough transform (VHT) to cater for arbitrary motion. Like the VHT, the new technique processes the whole image sequence, gathering global evidence of motion and structure. However, we do not assume constant linear velocity but rather allow arbitrary velocity. The method tries to find an optimal, smooth trajectory in the parameter space with maximum energy, where the latter incorporates both the structure of the moving object and the smoothness of motion. The constrained optimisation problem is solved using a temporal (time-delay) dynamic programming algorithm. Tracking in noise is much superior to the standard Hough transform.

Pronunciation by analogy (PbA) is a data-driven method for relating letters to sound, with potential application to next-generation text-to-speech systems. This paper extends previous work on PbA in several directions. First, we have included “full” pattern matching between input letter string and dictionary entries, as well as including lexical stress in letter-to-phoneme conversion. Second, we have extended the method to phoneme-to-letter conversion. Third, and most important, we have experimented with multiple, different strategies for scoring the candidate pronunciations. Individual scores for each strategy are obtained on the basis of rank and either multiplied or summed to produce a final, overall score. Five strategies have been studied and results obtained from all 31 possible combinations. The two combination methods perform comparably, with the product rule only very marginally superior to the sum rule. Nonparametric statistical analysis reveals that performance improves as more strategies are included in the combination: this trend is very highly significant (p < 0:0005). Accordingly for letter-to-phoneme conversion, best results are obtained when all five strategies are combined: word accuracy is raised to 65.5% relative to 61.7% for our best previous result and 63.0% for the best-performing single strategy. These improvements are very highly significant (p ∼ 0 and p < 0:00011 respectively). Similar results were found for phoneme-to-letter and letter-to-stress conversion, although the former was an easier problem for PbA than letter-to-phoneme conversion and the latter was harder. The main sources of error for the multistrategy approach are very similar to those for the best single strategy, and mostly involve vowel letters and phonemes.

Previous work has shown that connectionist learning systems can simulate important aspects of the categorization of speech sounds by human and animal listeners.Training is on representations of synthetic, exemplar voiced and unvoiced stop consonants passed through a computational model of the auditory periphery.In this work, we use the modern inductive inference technique of support vector machines (SVMs) as the learning system.Visualization of the SVM's weight vector reveals what has been learned about the voiced/unvoiced distinction.

A technique for recognizing planar objects in three-dimensional space is de-scribed. The object’s domain is not restricted to purely two-dimensional items but includes fairly flat real objects such as a pair of scissors. To tackle realistic matching problems, the proposed method is invariant to changes within a prescribed range of perspective viewpoints, scaling, shift and rea-sonable degree of occlusion. In brief, objects are recognized by a modified dynamic link matching. Our experiments show that the system is very suc-cessful in recognizing deformed objects due to perspective distortion, even in rather cluttered scenes. 1

Most prior works in neural networks for object matching focus largely on 2D problems only and often assume highly-constrained environments. In contrast, this paper takes a more generalized approach for planar object matching which is invariant to 3D perspective transformation and partial occlusion. The object's domain is not restricted to purely 2D items; it includes fairly flat real objects such as a pair of scissors. The proposed system uses multi-view model representations and objects are recognized by self-organized dynamic link matching. The merit of this approach is that it offers a compact framework for concurrent assessments of multiple match hypotheses by promoting competitions or co-operations among several local mappings of model and test image feature correspondences. A wide spectrum of rigorous test has been applied to the proposed system. Experimental results have demonstrated the system's ability to produce accurate matches even when the test image is cluttered with irrelevant features.

This paper describes a modification to, and a fast implementation of the Viterbi algorithm for use in stochastic letter-to-phoneme conversion. A straightforward (but unrealistic) implementation of the Viterbi (1967) algorithm has a linear time complexity with respect to the length of the letter string, but quadratic complexity if we additionally consider the number of letter-to-phoneme correspondences to be a variable determining the problem size. Since the number of correspondences can be large, processing time is long. If the correspondences are precompiled to a deterministic finite-state automaton to simplify the process of matching to determine state survivors, execution time is reduced by a large multiplicative factor. Speed-up is inferred indirectly since the straightforward implementation of Viterbi decoding is too slow for practical comparison, and ranges between about 200 and 4000 depending upon the number of letters processed and the particular correspondences employed in the transduction. Space complexity is increased linearly with respect to the number of states of the automaton. This work has implications for fast, efficient implementation of a variety of speech and language engineering systems.

Responses of both human and animal listeners to synthetic stop-consonant/vowel stimuli in which voice-onset time (VOT) is uniformly varied are known to be 'categorical' but an explanation of this phenomenon remains elusive. A 'composite' model consisting of a physiologically-realistic auditory model feeding its patterns of neural firing to an artificial neural network is described. When trained by (supervised) error back-propagation on the extreme, end-points of the VOT continuum, the composite model is capable of reproducing closely listeners' behaviour in classical categorical-perception (CP) studies. However, whether the model also reproduces the so-called boundary-shift phenomenon-whereby the phoneme boundary moves with place of articulation-apparently depends upon the precise details of the auditory model and so, by implication, upon subtle aspects of peripheral auditory processing. A first attempt at unsupervised training has been unsuccessful: the likely reason for this is outlined. It is anticipated that future work comparing the model's responses for unsupervised versus supervised training will help to elucidate the mechanisms of categorical perception.

We investigate empirically the performance under damage conditions of single- and multilayer perceptrons (MLP's), with various numbers of hidden units, in a representative pattern-recognition task. While some degree of graceful degradation was observed, the single-layer perceptron was considerably less fault tolerant than any of the multilayer perceptrons, including one with fewer adjustable weights. Our initial hypothesis that fault tolerance would be significantly improved for multilayer nets with larger numbers of hidden units proved incorrect. Indeed, there appeared to be a liability to having excess hidden units. A simple technique (called augmentation) is described, which was successful in translating excess hidden units into improved fault tolerance. Finally, our results were supported by applying singular value decomposition (SVD) analysis to the MLP's internal representations.

Fault tolerance is often mentioned as an important, intrinsic property of neural networks but it has not often been the subject of directed study. Such fault tolerance as does exist must depend strongly upon the exact nature of the internal representations captured during training, and the way these are distributed across the network. A representative pattern-recognition task is used to assess the fault tolerance of feedforward neural nets as a function of hidden-layer size. Damage resistance is found to increase with the number of hidden neurons, although this finding is sensitive to the exact performance metric employed. The technique of augmentation, which can increase the fault tolerance of a net of given size, is described. To understand the relations between the obtained measures of fault tolerance and the nets' internal representations, a number of analyses are used. These techniques have more usually been used to simplify network structure by identifying and removing redundancies, but they are equally applicable to the study of fault tolerance.>

An algorithm for inferring correspondences between letters and phonemes from a large set of word spellings and their associated phonemic forms is described. The algorithm uses two techniques to infer correspondences: delimiting and dynamic time warping (DTW). The first technique delimits the part of the word spelling and pronunciation that cannot be aligned with the existing set of correspondences. The second technique derives correspondences from the delimited part of that word. The inferred correspondences are evaluated in terms of translation performance tested with unseen words, proper names and novel words. The translation performance is compared with those obtained using the manually driven correspondences as the benchmark. Nonparametric statistical tests are used to establish whether the performances of inferred correspondences are significantly different from the manually derived correspondences.>

An investigation is made of the relation between the fault tolerance of a multilayer perceptron (MLP) and its redundancy as determined by the number of hidden-layer neurons (x). Damage was introduced by cutting connections. The application studied is the classification of coins according to their acoustic emissions after striking a hard object. Several MLPs were trained by backpropagation to discriminate acoustic emission data from 6 classes of coin. The nets had 259 input nodes, 6 output nodes, and x varying between 5 and 25. In addition, one single-layer network (x=0) was trained. Results show that the single-layer perceptron (SLP)-although able to classify the data with 100% accuracy under fault-free conditions-was far less damage-resistant than any of the MLPs.>

It is shown how syntactic neural networks can be applied to the problem of translating orthographic strings to phonetics strings, and vice versa, due to the symmetry of the model. This is unusual in text-phonetics translation systems; most systems have to be trained to operate in a single direction. Another novel feature is the lack of supervision required during training. The only requirement is that one have whole-word orthographic/phonetic symbol-string pairs. To test the system the authors have formed a lexicon of (6000, single-syllable) such pairs in English by extracting the relevant information from the machine-readable Oxford Advanced Learner's Dictionary. Results are presented for cases where the training set varies between 10 and 2000 words. In each case, the trained nets are tested on the training set and an equal-size (disjoint) test set. Present results are poor compared to conventional translation systems, but most of the errors may be due to the system's immaturity.>

An inference algorithm to obtain letter-phoneme correspondences for English using image-processing techniques is described. The main idea of the inference algorithm is to construct a table of association indices. The table is regarded as an image and image-processing algorithms are use to derive letter-phoneme correspondences. The advantages of the technique are that the association index has a theoretical basis in terms of modeling spelling and pronunciation as a homogeneous Markov chain, computational demand is low, and the inference process is not dependent on the ordering of the sample. The authors have implemented both a thresholding and a minimum spanning tree technique to derive correspondences. Illustrative examples are given for each technique.>

The authors believe that the lack of success of computational text-to-speech systems when presented with a previously unseen word is these systems' over-reliance on pronunciation rules. Therefore, the authors are developing a model of novel-word pronunciation based on analogy with known words which has its origins in psychological investigations of oral reading. A progress report is presented on implementing the model and previous results are extended by considering initial and final spaces as an integral part of the word. Various ways of handling these spaces in the analogy process are considered.>

On the basis of previous psychological research, a psychologically motivated approach to the technical problem of novel-word pronunciation within a text-to-speech system is presented. The psychological basis for, and the structure of, the resultant model are outlined. Initial results for a partial implementation are given.>

A syntactic neural network is equivalent to a parser for a certain type of grammar-in this case, strictly hierarchical context-free. This allows an efficient method for pattern description and has the added advantage of being a generative model. The authors show how the network itself can infer the grammar. Syntactic neural nets can model stochastic or nonstochastic grammars. The stochastic nets are properly probabilistic and are powerful discriminators; the nonstochastic nets are less powerful, but have straightforward silicon implementations with existing technology. Learning in syntactic nets may proceed supervised or unsupervised. In each case, the algorithm is the same; the difference lies in the data presented to the net. In prior publications, the authors applied syntactic neural nets to character recognition and cursive script recognition. The authors presently show that nonstochastic nets can perform signature verification with high reliability. This raises the possibility of signature verification on a robust smart card

A novel method of tracking the poles of the linear prediction (LP) model of speech is described. This uses a new, real (rs) plane in which z-plane pole-pairs are represented by a single point. The method reduces the frame-to-frame correlation permutations which need to be considered and avoids the problems which occur when Z-plane complex-conjugate pairs turn real. Divergence problems associated with iterative methods of root extraction are considered, as is the reduction of frame size by filtering of the LP coefficients.>

There is extensive evidence that the classical, stuck-at fault model, operating at the gate level, is inadequate for testing MOS VLSI circuits. By contrast, a ``nonclassical,'' switch-level model-directly representing open and short circuits in the interconnect and transistors stuck-open or stuck-on-allows important effects such as MOSFET bidirectionality and tristate behavior to be taken into account. This paper describes a new switch-level method for generating the singular cover of an MOS primitive gate using path algebras. The method yields tests to cover all specified interconnect open and short circuits, and all irredundant transistor stuck-open and stuck-on faults, should such tests exist. It relies on specification of an appropriate algebra by redefinition of the operators used in computing powers of a matrix. Two such algebras are required-one to generate tests for open circuit faults and the other for short circuit faults. Test generation for networks of primitive gates is achieved in two stages: after deriving singular covers for the primitives, a variant of the D-algorithm with modified ``D-drive'' is used. The approach is unified and powerful, having the potential to detect parasitic latch behavior and to generate tests for any of the current MOS VLSI technologies. In common with most present automatic test generation methods, it is restricted to combinational logic. Practical limits on the size of circuit which can be dealt with appear comparable to those set by use of the classical D-algorithm.

Automatic speech recognition is being increasingly employed for man-to-machine communication; and the physically disabled appear to be one class of users who could benefit particularly from the new technology. In this paper, the potential of speech recognition to aid disabled people is reviewed and assessed. Applications to mobility, robotics, environmental control and computer access are considered. Two specific applications, viz. environmental control and computerised text processing, are especially attractive and have formed the focus of extensive work at Southampton.