Recent studies have shown that nonlinear predictors can achieve about 2-3 dB improvement in speech prediction over conventional linear predictors. In this paper, we exploit the advantage of the nonlinear prediction capability of neural networks and apply it to the design of improved predictive speech coders. Our studies concentrate on the following three aspects: (a) the development of short-term (formant) and long-term (pitch) nonlinear predictive vector quantizers (b) the analysis of the output variance of the nonlinear predictive filter with respect to the input disturbance (c) the design of nonlinear predictive speech coders. The above studies have resulted in a fully vector-quantized, code-excited, nonlinear predictive speech coder. Performance evaluations and comparisons with linear predictive speech coding are presented. These tests have shown the applicability of nonlinear prediction in speech coding and the improvement in coding performance.>

Studies in the psychology of reading indicate that reading probably involves recognizing features which are present in letters, such as loops, turns and straight strokes. If this is the case it is likely that recognizing these features will be a useful technique for the machine recognition of cursive script. A new method of detecting the presence of these features in a cursive handwritten word is described. The method uses constrained snakes which adapt to fit the maxima in the distance transform of a word image while retaining their basic shape. When the shape has settled into a potential minimum its goodness-of-fit is used to determine whether a match has been found. The features located by this method are passed on to a neural network recognizer. Examples of the features recognized are shown, and results for word recognition for this method on a single-author database of scanned data with 825 word vocabulary are presented.>

International Computer Science Institute (ICSI), USA(Author list is alphabetical with the exception of the typist.)ABSTRACTThis paper describes the research underway for the ESPRITWERNICKE project. The project brings together a num-ber of different groups from Europe and the US and focuseson extending the state-of-the-art for hybrid hidden Markovmodel/connectionist approaches to large vocabulary, continu-ous speech recognition. Thispaper describes the specific goalsoftheresearchandpresentstheworkperformedtodate. Resultsare reported for the resource management talker-independentrecognition task. The paper concludes with a discussion of theprojected future work.Keywords: Recognition, Neural Nets, HMM.1. BACKGROUNDW

A model of dynamic complexity, a growing Gaussian radial basis function (GRBF) network, is developed by analyzing sequential learning in the function space. The criteria for adding a new basis function to the model are based on the angle formed between a new basis function and the existing basis functions and also on the prediction error. When a new basis function is not added the model parameters are adapted by the extended Kalman filter (EKF) algorithm. This model is similar to the resource allocating network (RAN) and hence this work provides an alternative interpretation to the RAN. An enhancement to the RAN is suggested where RAN is combined with EKF. The RAN and its variants are applied to the task of predicting the logistic map and the Mackey-Glass chaotic time-series, and the advantages of the enhanced model are demonstrated.>

It is shown how the compositional representation (CR) previously used for lexical access from sub-word recognizers for a relatively small word vocabulary can be extended to much larger vocabularies without further training. This is demonstrated for the DARPA Resource Management database where, using sub-word units as input, words are presented distributively over a fixed number of units and classified using a simple network. Initially, the architecture is trained on 147 words achieving an accuracy 91.2%. Then, leaving the recognizer unchanged, it is shown how additional output units can be added to the network to increase the vocabulary to the complete set of 975 phonetically distinct words. On this extended vocabulary the performance dropped to 66% but this drop is less than the expected drop due to the perplexity increase. Further improvement would be achieved by improving the performance on the original data set.>

Consideration is given to image filtering (temporal and spatial) in a neural system for transmitting images through a limited capacity channel, in the case of a noisy image at the receptors. The authors use an extension of Shannon's (1949) formula for the capacity of a Gaussian channel to determine the optimum filter to be used. For realistic image statistics, they show that the bandwidth of this filter is self-limiting, and it has a high frequency boost that disappears at low signal levels. This behavior is mirrored in biological retinas.>

Abstract A real‐time prototype speech recognizer has been implemented on a 66‐processor distributed‐memory parallel computer. Simple phrases are recognized in approximately 4 to 10 seconds. Scalability, performance and flexibility are the three main aims of this implementation, the ultimate goal being to construct a large vocabulary speech recognizer which responds quickly. A set of three techniques is investigated in this implementation: asynchronous methodology to minimize synchronization overheads, distributed control to avoid a central communications bottleneck, and dynamic load balancing to provide a flexible response to an unpredictable computational load. The effect on memory, processor time allocation and communications is observed in real‐time using hardware monitoring aids.

Four types of neural networks which have previously been established for speech recognition and tested on a small, seven-speaker, 100-sentence database are applied to the TIMIT database. The networks are a recurrent network phoneme recognizer, a modified Kanerva model morph recognizer, a compositional representation phoneme-to-word recognizer, and a modified Kanerva model morph-to-word recognizer. The major result is for the recurrent net, giving a phoneme recognition accuracy of 57% from the si and sx sentences. The Kanerva morph recognizer achieves 66.2% accuracy for a small subset of the sa and sx sentences. The results for the word recognizers are incomplete.>

A temporal reasoning system for solving practical planning problems is presented. The system is based on a nonmonotonic temporal logic which features homogeneous assertions, continuous time, and defeasible inferences. The model-theoretic semantics provide a basis for algorithms which generate inferences directly from axioms in the logic. The system's utility is demonstrated by a solution to the multiple agent collision avoidance problem. A formulation of this problem separates spatial and temporal considerations. The temporal reasoning system is used to predict collisions and generate efficient strategies for avoiding them.>

An analysis is made of the behavior of the Hopfield model as a content-addressable memory (CAM) and as a method of solving the traveling salesman problem (TSP). The analysis is based on the geometry of the subspace set up by the degenerate eigenvalues of the connection matrix. The dynamic equation is shown to be equivalent to a projection of the input vector onto this subspace. In the case of content-addressable memory, it is shown that spurious fixed points can occur at any corner of the hypercube that is on or near the subspace spanned by the memory vectors. Analysed is why the network can frequently converge to an invalid solution when applied to the traveling salesman problem energy function. With these expressions, the network can be made robust and can reliably solve the traveling salesman problem with tour sizes of 50 cities or more.

Connected utterance speech recognition systems using hidden Markov models have traditionally been based on the Viterbi algorithm, an instance of Bellman's principle. This is a two-pass algorithm: the first pass makes local decisions on the direction of potential optimal paths, and the second generates a globally optimal path from the local information. This study eliminates the necessity for a second pass by associating separate objects, representing the history of the recognition, with each recognition score. As a result, it is possible to impose an arbitrary syntax direction mechanism on the recognition by examining these history objects at each frame of the recognized speech and predicting possible recognition paths. This technique presents a natural arrangement for implementing speech recognizers on parallel computing architectures. A demonstration recognizer using a push-down automation to implement a context-free grammar on a transputer network is described.>

A connectionist structure for phoneme recognition in continuous speech is described. This has two main parts. The first is a sound subunit classifier in the form of a three-layer back propagation network which classifies speech subunits from frames of spectral speech data. This is followed by a sequence classifier in the form of a network of neural like-units which classifies phonemes from input sequences of subunits by their occurrence and duration. Results are given for a 15-phoneme subset of British English, for a single speaker. These include the difficult syllable initial and final stop consonants, fricatives, vowels, and diphthongs. The overall recognition accuracy achieved is 87%.>

Short term spectral analysis of source-filter modeling gives a parameterized description of the acoustic signal in terms of a sequence of vectors. These parameter vectors change slowly with time corresponding to a slowly moving vocal tract. The authors consider a model (temporal decomposition) that approximates the time variation by a set of target vectors and interpolation functions that overlap in time. They present a geometric interpretation of the approach, describe an algorithm for decomposing a given utterance into parameters of such a model, and discuss how such modeling can be used in speech recognition systems.>

An integrated system which is part of a larger project in robotic assembly is described in terms first of its separate components and then of its integrated form and performance. The parts to be assembled are previously designed by CAD, and therefore geometrical models of the parts are available in a solid modeler which is incorporated in the system. In the integrated system presented, simplified versions of the component parts of an aircraft equipment tray are recognized and located by vision from an arbitrary initial state and moved to a final state appropriate to assembly. The components of the system described are: low-level vision processing using a transputer array; an intelligent vision system integrated with a solid modeler for part recognition and location; a path planner using 2/sup n/-array tree search; and the control of a PUMA robot. The integration of the components and the performance of the prototype system are described.>

Most forms of hidden Markov models (HMMs) do not incorporate contextual information and consequently may have problems in distinguishing isolated errors in utterances from significant features of short duration. A HMM has been derived, which treats sequence observations as a whole and not as a series of independent random variables. The model is based upon a distortion measure, which compares observations against templates contained in the model. From a sequence of observations, a typical distortion is computed and this is used to match a test utterance to the model, the utterance being regarded as a string of observation sequences corresponding to state occupancies. In addition, the length of a sequence is modelled explicitly. The natures of the duration model and distortion measure are left undefined in order to accommodate alternative parameterisations of the speech. In the training algorithm, the task of re-estimating the parameters is reduced to simple subsidiary problems.>

A probabilistic algorithm is described for the estimation of the lexical stress pattern of English words from the acoustic signal using hidden Markov models (HMMs) with continuous asymmetric Gaussian probability density functions. Adopting a binary stressed-unstressed syllable classification strategy two five-state HMMs of the left-to-right type were generated, one for each stress value. Training observation vectors were extracted from a corpus of bisyllabic stress-minimal word pairs and consisted of nine acoustic measurements based on fundamental frequency, syllabic energy and coarse linear prediction spectra. Evaluation of both models using a set of recordings of the same word pairs yielded an average stress recognition rate of 94%.>

Multipulse-excitation has greatly improved the speech quality achievable from linear predictive coders which previously required speech to be classified as voiced or unvoiced for excitation purposes. Multipulse removes the need for voicing classification, it improves speech quality by enhancing the excitation and offsets errors in the vocal tract filter. The authors have investigated a multipulse-excited channel vocoder. The prime objective being to improve the speech quality of the channel vocoder and achieve multipulse linear prediction quality. We describe how a particular channel vocoder is incorporated into a multipulse algorithm, and report the initial results of our multipulse-excited channel vocoder investigation.

A hybrid pattern recognition machine is presented for the classification of place and manner of articulation categories in continuous speech. Qualitative speech knowledge is represented declaratively as an augmented context-free grammar. This representation is supported at each category by probabilistic classifiers which enable an underlying quantitative description to be learned from examples of training data. Motivating factors for feature extraction and choice of category and structure in phonetic IKBS are discussed.

There has been much work in using Hidden Markov Models to model different types of linguistically defined units such as words, syllables and phonetic-type units. Here we look at the problem from the other direction and try to use the states obtained from a Markov model to find our own linguistic units. We look at the problem at two levels, the first at the sub-word level to find significant segment labels and the second at the grammar level in an attempt to deduce the grammatical units of a given vocabulary from the emission probabilities of a Hidden Markov Model.

A handwritten script recognition system is presented which uses Hidden Markov Models (HMM), a technique widely used in speech recognition. The script is encoded as templates in the form of a sequence of quantised inclination angles of short equal length vectors together with some additional features. A HMM is created for each written word from a set of training data. Incoming templates are recognised by calculating which model has the highest probability for producing that template. The task chosen to test the system is that of handwritten word recognition, where the words are digits written by one person. Results are given which show that HMMs provide a versatile pattern matching tool suitable for some image processing tasks as well as speech processing problems.

A new extension to articulatory modeling is described which modifies the areas as a function of the flow. The model is described by ten parameters which determine the longitudinal and vertical positions of the teeth, tongue body and lips. The use of continuous spatial derivatives of the area simplifies the mathematical analysis by producing a simplified set of equations to be solved iteratively, and the inclusion of nonparallel flow in the areas around constrictions produces a more accurate model. Results show that the model produces close matches to reference spectra with little error in formant frequency and 3-6 dB error in bandwidth for continuous vowels.

Both in Europe and the USA there is a significant growth in the availability and use of videotext services. Videotext covers both viewdata and teletext. Teletext is in far greater use than viewdata, particularly in Britain and the rest of Europe. With the growth in the use of personal computers, the possibility is arising to combine the data capacity of teletext with the flexibility of computers. One option available is to provide speech output. The system described here is a research study for converting teletext into speech. The study examined constrained text - the teletext weather information. This enabled most of the research effort to be concentrated on the synthesis sections. Figure 1 gives an overall view of the system. Modularity was of prime concern to allow flexibility in altering the composition of each section. Most of the software is written in C, with the parser written in BCPL. The system runs on a DEC PDP 11/60.

A new speech training aid providing visual feedback for deaf children is described. The Fricative and Timing Aid (FTA) distinguisnes between silence, voiced sounds and unvoiced sounds and displays their presence and duration within words or phrases. It provides target and attempt traces, allowing subjects to perceive and correct errors in speech. The FTA was used under controlled conditions to teach fricative production and voicing contrasts, over six months. The results show that (i) for children learning voicing contrasts, those using the FTA made more progress than those not using it and (ii) for children learning fricatives the progress made by FTA users and non-users was similar. However, the FTA made speech training easier and more efficient.

A particularly demanding application of pitch detection is in speech training of the deaf. The prevalance of defects such as breathy, creaky or hoarse voice means that pitch detectors are prone to errors. To overcome this an error correction algorithm has been developed for the Dubnowski, Schafer and Rabiner (D.S.R.) pitch detector. The resulting pitch detector overcomes these problems. It is a real-time procedure which can be implemented in hardware or software. Its performance is comparable with the nonreal-time SIFT algorithm and may thus have wider applications.

There is interest in speech research for a model that captures all the relevant features of the intonation contour. Such a model is capable of generating natural intonation contours from the minimum amount of input data. Much of the work on this type of modelling has been performed for Dutch (1). Their model gave a result that was "perceptually-equivalent" to an original contour. A deficiency in this type of description is that perceptually-equivalent is not synonymous with natural. Although a contour may be perceived as being equivalent to an original, it may not necessarily be indistinguishable in an auditory sense from an original intonation contour. The model described in this paper has the capacity to produce intonation contours for English that are perceptually-equivalent and also auditorily indistinguishable from the original. The model consists of the concatenation of contour segments; each segment being modelled by a quadratic equation.

This paper describes an experimental system for synthesising speech via a computer model of the vocal tract. The system has been used to generate /VCV/ syllables and diphthongs. The vocal tract model is a 'ball in the mouth' simulation that calculates formants for each setting of the seven control parameters. Computationally, it consists of a set of polynomials that can be evaluated efficiently at run time. The more lengthy calculations needed to derive geometrical shapes, area functions and then resonances for various articulatory settings are performed off line, as part of the analysis procedure that derives the polynomial formulae.

A transformation is described which converts the vocal tract areas derived from the linear prediction model to vocal tract widths. The transformation is described by the equationw = c(kA)^{p}wherew= vocal tract widthA= vocal tract areac= constantp= power constant andk= absolute areas criterion. Optimum values ofcare given and the best criterion for making the linear prediction areas absolute was that the maximum area is a constant. The transform has been implemented in a real-time articulatory display which gives an overhead of only 14% over simple linear prediction area extraction. The display was found to be capable of showing the position of any vowel. Potential applications include its use as an aid to articulator placement in deaf speech training.

Summary form only given. The paper gives a solution of the problem of separating n speakers observed by m microphons by modelling the n speakers as a multi-input, multi-output (MIMO) system. The paper demonstrates that for separation there is the requirement m = n, not surprisingly, shows the limitations of machine analysis in the cocktail party situation m = 2, gives an algorithm with results for the example m = n = 2 and comments on some applications of the method in conference telephony, speaker position measurement etc.

A short-time dyadic autocorrelation function (dacf) and a short-time Walsh energy spectrum of the first kind are defined in the Walsh-Fourier domain. The "natural" choice of the short-time functions does not lead to a Walsh-Fourier transform pair (dyadic Wiener-Khintchine theorem), and thus a second kind of short-time dacf and short-time Walsh energy spectrum are defined as the Walsh-Fourier transforms of the first kind. This leads to a meaningful and convenient Walsh transform pair between the first short-time Walsh energy spectrum and the second short-time dacf. The measurement procedures for both kinds of functions are discussed, and the mean values of these short-time functions are shown to be related to the corresponding long-time functions.