The Independent Component Analysis (ICA) technique is amenable to a coarse-grain parallel-processing chip architecture. However, the computation of nonlinear functions is critical in this algorithm. An efficient hardware approach is presented here for the computation of such functions , some of which are compound and concatenated. All of the needed functions are regularized into a single algorithm so a new result produced on each cycle even if the function changes from one cycle to the next. The worst case arithmetic error is predicted and bounded. This enables the designer to quickly select the architectural parameters without expensive simulations, while insuring the desired accuracy. A design is presented for the 32 bit fixed point case.

This work is motivated by the desire to map the independent component analysis (ICA) technique to a coarse-grain parallel-processing chip architecture. As in many other advanced DSP algorithms, the computation of nonlinear functions is critical in this algorithm. We discuss an efficient hardware approach to the computation of such functions for the ICA, some of which are compound and concatenated functions. All of the needed functions are regularized into a single efficient algorithm, and a new result is produced every cycle - in a pipelined mode - even for a different function every cycle. The underlying principle which makes the combined goals of high-speed and multi-functionality possible is significance-based polynomial interpolation of ROM tables. Very importantly, the paper uses a key formula for predicting and bounding the worst case arithmetic error. This theoretical result enables the designer to quickly select the architectural parameters without the expensive simulations, while guaranteeing the desired accuracy

A 3D heterogeneous sensor using a stacked chip has recently been proposed. While the sensors are located on one of the planes, the other planes provide for analog processing, digital signal processing, and wireless communication. This paper focuses on its DSP plane, in particular on the implementation of the ICA (independent component analysis) algorithm in the DSP plane. ICA is a recently proposed method for solving the blind source separation problem. The objective is to recover the unobserved source signals from the observed mixtures without the knowledge of the mixing coefficients. We present a parallel architecture utilizing the reconfigurable J-platform, which employs coarse-gain VLSI cells. These include a universal nonlinear (UNL) cell, an extended multiply accumulate (MA PLUS) cell, and a data-fabric (DF) cell. The coarse-grain approach has the distinct advantages of reduced external interconnect, much reduced design time, and manageable testability. Additionally, the other algorithms needed for the 3D HSoC can also be mapped on to the same resources, by time multiplexing, thereby reducing the silicon area needed.

This paper considers defect tolerance issues for parallel computing systems based on a new interconnection network, namely "Tori connected mESHes (TESH)". Key features of this network are the following: it is hierarchical, thus allowing exploitation of computation locality and systematic expansion up to a million processors; and it appears to be well-suited for VLSI/ULSI realization, including 3D implementation. The goal here is to present efficient reconfiguration algorithms for such hierarchical parallel computing systems. Despite the dramatic improvement in defect density in recent years, it is still necessary to provide redundancy and defect circumvention to achieve acceptable system-level yields for large multicomputer systems. The TESH-based parallel systems are no exception. Therefore, we develop placement and routing algorithms that assign logical nodes to healthy physical nodes and configure switches to bypass the defective cells, switches and links. Simulations indicate that the placement is nearly 100 percent effective, while the routing performance diminishes with increasing defect density for a given extent of redundancy. The approach scales up well because, in TESH networks, essentially the same kind of sparing is used at all levels.

This paper develops a signaling system for high bit rate wireless LANs made possible by a new class of signaling pulses. Called OWSS pulses, they are generated through a combination of orthogonal wavelet-division multiplexing (OWDM) and spread-spectrum concepts. Specifically, they are derived from the OWDM pulses by spreading them in the wavelet domain, using a suitable family of PN codes. Unlike the pulses used in OFDM, TDMA, and even OWDM, these new pulses have both a wide time support and a wide frequency support. Some of the advantages realized are: (a) single or multiple user capability, (b) inherent frequency diversity for overcoming frequency selective fading, (c) continuously pipelined operation (in contrast to OFDM), and (d) a high bandwidth (BW) efficiency, which is about twice that for DS-CDMA (assuming rectangular chips for DS-CDMA). The target bit rates for this work range from 100 Mbps upwards, in the 5.7 GHz band.

This paper discusses very large scale integration (VLSI) issues, including reconfiguration and yield, for a new interconnection network, "Tori connected mESHes (TESH)". Its key features are the following: (1) it is hierarchical, thus allowing exploitation of computation locality as well as easy expansion up to a million processors, (2) it permits efficient VLSI/ULSI (ultralarge scale integration) realization, and (3) it appears to be well suited for three-dimensional (3-D) implementation. This is due in part to the fact that it requires far fewer number of vertical wires than most other multicomputer networks of comparable diameter, as demonstrated by a 4096 node example. Presented in the paper are the architecture of the new network, node addressing and message routing, VLSI considerations, and most importantly, the reconfiguration and yield studies.

A unified approach to the design of linear- and nonlinear-phase QMFs is developed. Formulated as an optimization problem, the design procedure is shown to translate into an eigenvalue-eigenvector problem. To find the optimal filter an algorithm is presented, which typically converges in a few tens of iterations. The flexibility of our design procedure permits several practical extensions to be made readily. These are (a) inclusion of frequency-weighted stop band energy criteria, and (b) inclusion of finite word-length constraints. We have successfully used our filters to applications such as image coding and analysis; here, their use in wavelet-series analysis of oceanographic data is demonstrated.

A neural network based motion-estimation technique is developed that is applicable to purely translational and as well as affine movements. It produces significantly lower motion compensated frame differences than the existing approaches, such as the logarithmic block matching and full search algorithms. This advantage, which is particularly dramatic when new and uncovered background is introduced into the image or new objects are formed, arises because a weighted combination of candidate macroblocks is used for reconstruction formulated in terms of a modified Hopfield neural network, the procedure consists of two stages: estimation of the neural network parameters, followed by estimation of affinities for the candidate macroblocks.

The development of generic cells to perform linear and nonlinear arithmetic is critical to cost-effective rapid prototyping. Our LARP architecture deploys generic cells, thereby offering the capability of mapping a wide variety of signal and image processing algorithms on to the same pool of resources. This paper presents a further advance in this direction. Specifically, it reports on the development of a complex-argument nonlinear universal cell, which achieves both high speed and high accuracy. To accomplish this, low-degree MIP polynomials are used, as explained in the paper. Four specific functions, namely the reciprocal, absolute value, logarithm, and the square-root, are considered, but the procedure is extendible to other functions as well. A 24and a 16-b chip are described. Finally, their application to an antenna array processor is presented which is amenable to large-area-device implementation.

Abstract only given. This paper presents a wavelet based image coding method achieving high levels of compression. Our eigen-designed quadrature mirror filters are implemented in a multi-resolution subband decomposition system. Symmetrical extension and windowing of the multi-scaled subbands are incorporated to minimize the boundary effects. Next, the embedded zerotree wavelet (EZW) coding algorithm is used for data compaction. In our experiments, the EZW coded stream can be further reduced to yield a smaller symbol string by not coding the zero symbol in the last three subbands of the wavelet map, i.e. HL1, LH1, and HH1 which leads to an improved EZW algorithm. Additional compression is obtained with an adaptive arithmetic coder. Results of coding Lena are tabulated. For a low PSNR of 27.69 dB at a bit rate of 0.8554 for the aerial image, the quality of the decoded image is visually acceptable. No blocky effects are evidenced as compared to the JPEG coding technique.

Advances in artificial neural networks have led to several applications such as in computer-vision, image restoration, speech recognition, and pattern classification. However, for widespread practical use these and many other potential applications require large, high-speed networks implemented in efficient custom hardware. This paper investigates an important issue in the implementation of such networks, namely the avoidance of multipliers. Secondly, it evaluates their application in digital video, specifically in high-performance motion prediction and frame reconstruction. The authors discuss the multilayer perceptron and Hopfield neural networks in particular.

A novel vector subbands/ECVQ approach for image coding is presented which incorporates the concept of activity map. A key issue in vector subband coding is the efficient compression of the small energy-albeit perceptually significant, upper-bands of the decomposed images. As the components of the significant vectors in the upper subbands display the contour information of the original images, therefore, the locations and orientations of such (predominantly) edge information can be determined from the original image. The availability of this knowledge motivated us to generate the activity map, which in turn is used to obtain transmission control maps. The control maps are used to turn on/off the transmission of vectors. Most importantly, the approach is extended and applied to color image coding. Experimental results show remarkably good performance of the proposed approach.

Digital television seeks to reduce the bit rate through the elimination of interframe redundancy. The more accurate the motion estimation, the lower the encoded bit rate with the same picture quality, resulting in more channels the provider can deliver to the consumer at a reduced price. We develop a neural network motion estimation technique which produces 4 to 11 dB improvement of peak signal to noise ratio (PSNR) over the other methods, thus creating the potential for improved compression. In contradistinction to the other methods, there is no need to carry out fractional-pixel interpolation. The technique is applicable to purely translational, shear, and zoom sequences. Further, it is formative in that it can form objects that are formed in the present frame and are non-existent in the previous frame.

High-speed coprocessors for computing nonlinear functions are important for advanced scientific computing as well as real-time image processing. In this paper we develop an efficient interpolative approach to such coprocessors. Performed on suitable subintervals of the range of interest, our interpolation which uses third degree polynomial is adequate for many elementary functions of interest with double precision mantissas. Our method requires only one major multiplication, two minor multiplications and a few additions. The minor multiplications are for the second and third degree terms, and their significant bits are much fewer than those of the first degree term. This leads to a very fast and efficient VLSI architecture for such coprocessors. It appears that polynomial based interpolation can yield considerable benefits over previously used approaches, when execution time and silicon area are considered. Further, it is possible to combine the computation of multiple functions on a single chip, with most of the resources on the chip shared for several functions.>

A neural network based motion-estimation technique (for successive video frames) is developed that is applicable to purely translational as well as affine movements. In addition to yielding highly accurate results, a benefit of this technique is the elimination of fractional-pixel interpolation and corresponding search. The technique is formulated in terms of a modified Hopfield neural network. The procedure consists of two stages: estimation of the neural network parameters from the present and previous frames, followed by the estimation of the affinity weights for the candidate macroblocks. The reconstructed macroblock then consists of the sum of the candidate macroblocks weighted by their affinity values as defined. Experiments on translational motion indicate that this technique produces a 5 dB higher peak signal to noise ratio (PSNR) than the full search algorithm, 6 dB higher than the logarithmic block matching algorithms, and 2 dB higher than some of the very recent motion estimation algorithms. Preliminary experiments using the neural network affine algorithm on synthetically sheared images produce an 11 dB higher PSNR than the full search and LBM. Further, due to the neural network's fault tolerant nature and inherent parallelism, potential exists for VLSI implementation as an interconnection of small cells on a single chip.

A technique is proposed for the detection of tumors in digital mammography. Detection is performed in two steps: segmentation and classification. In segmentation, regions of interest are first extracted from the images by adaptive thresholding. A further reliable segmentation is achieved by a modified Markov random field (MRF) model-based method. In classification, the MRF segmented regions are classified into suspicious and normal by a fuzzy binary decision tree based on a series of radiographic, density-related features. A set of normal (50) and abnormal (45) screen/film mammograms were tested. The latter contained 48 biopsy proven, malignant masses of various types and subtlety. The detection accuracy of the algorithm was evaluated by means of a free response receiver operating characteristic curve which shows the relationship between the detection of true positive masses and the number of false positive alarms per image. The results indicated that a 90% sensitivity can be achieved in the detection of different types of masses at the expense of two falsely detected signals per image. The algorithm was notably successful in the detection of minimal cancers manifested by masses </=10 mm in size. For the 16 such cases in the authors' dataset, a 94% sensitivity was observed with 1.5 false alarms per image. An extensive study of the effects of the algorithm's parameters on its sensitivity and specificity was also performed in order to optimize the method for a clinical, observer performance study.

This paper discusses high-speed array implementations of two image processing algorithms, namely the 'Hough transform for detection of line segments', and 'backprojection in CT image reconstruction'. A multi-chip-module (MCM) construction is proposed consisting of three types of chips, a high speed multi-function nonlinear chip, a flexible multiply-accumulate chip, and an image kernel chip. Called V-array, it can be configured to have eight Hough modules, so as to produce the Hough transform of a 1024/spl times/1024 image in an estimated 13 ms in 2.0 micron CMOS technology (6.6 ms in 1.0 micron CMOS technology). Similarly, a V-array MCM can accommodate eight CT modules, which can produce the backprojected image in 209 ms in 2.0 micron CMOS technology (105 ms in 1.0 micron CMOS technology). To gain a significant speed advantage, we have developed an advanced multi-function cell for performing any one of four nonlinear operations: square-root, reciprocal, sine/cosine, and arctangent-all realized in a single chip, accessible on a selectable basis. A 16 bit four-function "one cycle" VLSI chip, fabricated in 2.0 micron CMOS technology, is presently available which outputs a new result every clock cycle. Using this nonlinear cell and two other cells, an application level Hough transform module and a CT module are presented.>

This paper discusses an efficient interpolation method for nonlinear function generation. Based on this, a 24 bit VLSI cell, capable of computing the (1) square root, (2) reciprocal, (3) sine/cosine, and (4) arctangent functions, is presented for single precision floating-point applications. A 53 bit version, suitable for double precision computations, is also presented. Finally, an extension of the method to the two dimensional nonlinear functions is briefly addressed.>

A neural network based motion-estimation technique is developed, that is applicable to sub-pixel as well as large movements. Experiments on typical test frame sequences indicate marked improvement in accuracy of motion vector estimates over the MPEG logarithmic block matching algorithm. The method utilizes a modified Hopfield neural network. Due to the neural network's fault-tolerant nature and parallel computation capability, fast, accurate, and reliable results are obtained. Application to MPEG based video compression is also discussed.>

This paper presents a novel technique for motion estimation in video frame sequences. It uses a modified Hopfield neural network. The procedure consists of two stages: estimation of the neural network parameters from the present and past frames or subimages, followed by estimation of the motion vector. The latter utilizes a dynamic iterative algorithm to minimize the energy function of the neural network. Due to the neural network's fault-tolerant nature and parallel computation capability, fast, accurate, and reliable results are obtained. The usefulness and accuracy of the approach is demonstrated upon both synthetic and real images.>

The arithmetic error of a fast reciprocal 16-b cell is analyzed. This VLSI cell computes the result in two clock cycles by the use of a very small ROM table and innovative second-order interpolation. The reliability of the predictive formulas (for the arithmetic errors of the cell) is demonstrated by comparing the predictions with computer simulation results. The methodology outlined can easily be extended to other VLSI cells.>

A novel high-speed cell, capable of performing a square-root or a reciprocal function in two clock cycles, is presented. Its performance signifies an estimated three-fold increase in speed over existing approaches. Furthermore, since both functions are performed on the same cell, an area advantage of a factor of two is realized. The cell therefore has a figure of merit of six. The underlying principle used is second order interpolation which leads simultaneously to high accuracy and a very small ROM table. The statistical performance of the cell is presented. The application of the new cell is demonstrated by its use in two systolic architectures, namely an L-U decomposition architecture and a Cholesky decomposition architecture.>

A methodology is presented for the statistical analysis of the arithmetic error of VLSI architectures. Attention is focused on predicting the arithmetic errors of the overall architecture as a function of register lengths, the type of word representations used, and the precision of the arithmetic blocks (adder/subtracters, multipliers, shifters, dividers, etc.). The methodology is demonstrated by application to a newly developed for use in a large-scale signal processing wafer architecture, a L-U decomposition array. The wafer design, intended for radar signal processing, is part of an ongoing effort on restructurable wafer-scale integration for large-scale signal processing.>

A systolic architecture for L-U decomposition using a recently developed reciprocal cell is presented. The arithmetic of this new cell is based on second-order polynomial interpolation, which results in high speed and inherent stability in inversion. In part, the speed of the cell is realized by use of a special empirical mapping. For the 16-b mantissa of a floating point number, the lookup table size required is 64 words, and the RMS value of the error is approximately one-third of the LSB of the 16-b result.>

The CCITT recommendation, G.721 for a 64 kbps PCM to 32 kbps ADPCM transcoder algorithm does not easily lend itself to implementation on general purpose digital signal processors. The authors present a bit-for-bit compatible realization of the algorithm and a functional real time test bed for algorithm modifications. The hardware of the unit, USF32-2 is built around a single TMS320C25 microprocessor and has added circuitry to decrease computation time and facilitate testing. The base algorithm and enhanced version can be selectively compared. The encoder/decoder requires a total of 1301 clock cycles, but can be compressed to 1250, thereby enabling one to run both the encoder and the decoder on a single TMS320C25 processor.>

For noisy signals it is not enough to achieve a match between the first p+1 correlations of the measured data and that of the AR model. The higher order correlations begin to diverge as the level of data noise increases. In an effort to perform extended correlation matching we have found the need to use an Autocorrelation Distortion Function (ADF), defined as the difference between the sample acf of the noisy-data and the statistical acf of the true signal. Using the model bα|n|for the ADF we develop a procedure for improved AR modeling of a stationary time series. In this procedure we estimate the parameters of the ADF in such a way that while the autocorrelation of the signal model plus noise matches the data acf exactly for the first p+1 lags, simultaneously, the mismatch at certain higher-order lags is minimized. These higher-order lags are taken to be p+1, .... ,p+q or, alternatively, they are chosen according to a peak-picking scheme described in the paper.

In this paper we discuss the Half Fourier Transform (HFT) and explore its application to radar-return signals with specular components. It is shown that this transform enables separation of the desired part from the specular impulsive components. The effectiveness of this technique and the computer program developed, is demonstrated by simulation examples. Our program uses an optimization package which performs minimization of a nonlinear sum-of-squares functional with a model trust region strategy.

The CCITT has recently proposed a Draft Recommendation for a tandemable 64 kbps PCM to 32 kbps ADPCM transcoder algorithm [1],[2]. This paper presents a real-time hardware and software realization of the algorithm based on a single TI TMS 32010 each for the encoder and decoder. Objective and subjective results indicate that our changes in the adaptive quantizer and prediction filter, which make this single µP chip realization possible, do not degrade system performance relative to the CCITT algorithm.

Contamination of speech, for example by environmental noise, is sometimes unavoidable. Under such circumstances the familiar LPC analysis technique, either for low bit-rate coding or for automated recognition at the receiver, becomes fragile thus jeopardizing the system objective. In this paper we present an extended correlation matching approach for LPC analysis which results in good spectral matching between the true speech spectrum and the all-pole model spectrum, especially for the first three formants. The method has been tested on both synthetic as well as real speech and has been found to yield satisfactory results. Informal listening tests confirm the success of the technique developed.

This paper derives a new form of discrete Kalman estimator for speech parameter estimation. It is developed by considering the parameters themselves to be a process, in a manner similar to that of Lee and Athans. In principle, the algorithm is a two-tiered estimation procedure, implemented via an augmented parameter vector. However, the effective dimensionality of the estimator, and hence the transmission bit-rate, is not changed since the augmentation vector is not transmitted. Improved tracking of speech formants is achieved. Furthermore, this tracking ability is maintained even at very low levels of the speech signal, allowing better estimation during the latter stages of the closed glottis interval. Experimental results are presented for a synthetic "diphthong" with linearly time-varying formants, and a segment of real speech.

This paper presents a recursive approach to the estimation of pulses in speech excitation. The detector is derived from the generalized likelihood ratio method but does not assume that the system parameters are known exactly. It utilizes some of the characteristics of the speech signal to determine the a-priori probability of the primary pulse in an excitation sequence and estimates the pulse interval and amplitude envelope. This algorithm is particularly useful in conjunction with recursive or pitch-synchronous analysis methods. It not only provides a means for estimating signal features such as pitch, but also for synchronizing and adapting analysis methods with changing excitation patterns. Examples are presented for a synthetic vowel with linearly time-varying pitch and a segment of real speech.

There are many approaches to the computation of the auto-regressive model of a signal. To gain familarity with the tradeoffs of various methods, the performances of three popular algorithms are examined in this study. The performances of the covariance (COV), the singular-value-decomposition (SVD), and the pencil-of-functions (POF) methods are compared. The accuracy of the pole locations for the modeled signal is measured as a function of the signal-to-noise ratio (SNR) and the record length. The results of this study indicate that for an SNR below 50 db, both the SVD and the POF are superior to the COV method. In terms of error only, the SVD is either comparable or slightly superior to POF. However POF, with noise-correction incorporated, may prove to be comparable to SVD.

Recently, modified linear prediction (MLPC) was proposed as an alternative to conventional linear prediction. In MLPC the nth sample is predicted in terms of certain filtered signals. It is shown that this technique can be quite effective for accurate estimation of signal poles, provided that the signal energy is concentrated in a small portion of the frequency band. Performance comparable to other high accuracy techniques, such as the SVD, can be achieved with reduced computations.

The introduction of multi-pulse excitation for LPC coders has increased the quality achievable for digitally coded speech at bit rates in the 9.6 Kbs range. A simplified multi-pulse analysis is proposed here with particular emphasis on the speech model developed as a result of a two-pass method. In the first pass, estimated LPC parameters generated by conventional covariance analysis are used to generate the forward prediction error; the multi-pulse sequence is then detected by thresholding the residual. The second pass generates the final LPC parameters by a covariance analysis incorporating knowledge of the estimated pulse locations and amplitudes along with perceptual synthesis error weighting. Experimental results are presented to demonstrate the method.

A novel technique for the design of analysis/synthesis filter banks, such as quadrature-mirror filter (QMF) banks, is presented. The technique is iterative, but each iteration step is mathematically simple. It is based upon a sectional approximation of the filter bank wherein the synthesis section is taken to be known from the previous iteration, and the analysis section is optimized according to the performance criterion. The strategy is then reversed, and the synthesis section optimized. At each iteration, each of these two steps results in an eigenvalue-eigenvector problem of a matrix with dimensionality equal to the number of taps. The technique is exemplified in detail for the QMF bank design problem.

In this paper we discuss the use of a time-weighted error criterion in recursive speech parameter estimation. We show that the ability of Kalman type estimators to track time-varying parameters is enhanced by using a time-weighted error criterion. The modification to the estimator is easy to implement and well-suited to adaptation. We demonstrate the performance of the modified estimator on both synthetic and real speech.

From a single transient test, we determine all of the transfer function parameters of a test system, such as a filter, a communication channel, or a loudspeaker. The transfer function is obtained from the time-domain test data (recorded input and output) using a digital signal processing technique, specifically the pencil-of-functions technique. The methodology is demonstrated both by a laboratory test case and a simulation example.

Resolution of two sinusoidal signals of frequencies f1and f2requires a time record of length at least1/|f_{1}-f_{2}|. Resolution of two decaying exponentials of complex frequencies s1= σ1+jω1and s2= σ2+jω2, with\sigma_{2} \leq \sigma_{1}, by correlation coefficient is independent of the ratios σ2/σ1, and (ω1- ω2)/σ1for a time record of length greater than3/|\sigma_{1}|. This is also the condition for near-orthogonalization of a set of complex exponentials, with small error.

In autoregressive coding of speech, it is essential that the LPC model contain accurate information about the first three formants; specifically, that the LPC spectrum should reproduce the correct formant frequencies and the corresponding bandwidths. We shall demonstrate that the modified linear predictive coder (MLPC) is superior to the widely used linear predictive coder (LPC) when the data frames are short. MLPC uses a new error criterion. Indeed, prediction is sought in terms of certain filtered signals obtained by passing the speech into a cascade of first order digital filters. The normal equations for this modified linear predictor are derived and a reversible window, suitable for this new formulation, is proposed. Application to speech is demonstrated.

From a single transient test we determine all of the parameters of a vented-box loudspeaker system. A complete electroacoustical circuit representation is thus obtained from the time-domain test using digital signal processing techniques. First, we apply the pencil-of-functions technique [5] to the digitally recorded response and determine the impedance-function coefficients of the circuit. Next, using the relationships between the impedance coefficients and the circuit parameters, we show that every circuit parameter can be computed without resorting to simultaneous nonlinear equations.

The paper describes very recent and promising results on a new approach to noise correction for pole-zero modeling of signals by the pencil-of-functions method. In contradistinction to spectral subtraction techniques where only amplitude information is emphasized (and phase is ignored), we propose a method that (a) estimates the noise spectral density for the frame, and then (b) performs the subtraction of the noise correlation matrix from the correlation matrix of the information signals. The information signals are, of course, generated by processing the test signal by a cascade of first-order low pass signals, as described in earlier papers. Improvement resulting from this noise correction technique is particularly remarkable when judged by the pole-estimation accuracy, A comparison of the poles for multiple noise runs for three methods, namely, (a) Prony method, (b) Pencil-of-functions method, and (c) Pencil-of-functions method with noise correction, is given. The signal considered was the impulse response of a third order rational filter H(z), masked by noise with an SNR=12 dB.

In the conventional linear predictor the nth sample is predicted as a linear combination of the previous p samples. In this paper we seek the prediction of the nth sample as a linear combination of certain filtered signals. The normal equations for this modified form of predictor are derived and a window function, suitable for this new formulation, is proposed. For brevity we will consider only deterministic signals. The advantages of this modified linear predictor are demonstrated by an example.

In a recent paper a time-domain method for the measurement of the parameters of a loudspeaker-driver was described. Here we extend the technique to the more interesting problem of measuring the parameters of a vented-box loudspeaker system. We use the pencil-of-functions technique, and show that all of the loudspeaker system parameters can be determined straightforwardly from a single time-domain test. Results of computer simulation, demonstrating the effectiveness of the two-step approach, namely, identification of the impedance function by pencil-of-functions method and therefrom the computation of the physical parameters, is given in the paper. Results of experimental studies are expected to follow in the future.

Pole-zero modeling of low-pass signals, such as an electromagnetic-scatterer response, is considered in this paper. It is shown by use of pencil-of-functions theorem that (a) the true parameters can be recovered in the ideal case (where the signal is the impulse response of a rational function H(z)), and (b) the parameters are optimal in the generalized least-squares sense when the observed data are corrupted by additive noise or by systematic error. Although the computations are more involved than in all-pole modeling, they are considerably less than those required in iterative schemes of pole-zero modeling. The advantages of the method are demonstrated by simulation example and through application to the electromagnetic response ofa scatterer. The paper also includes very recent and tantalizing results on a new approach to noise correction. In contradistinction with spectral subtraction techniques, where only amplitude information is emphasized (and phase is ignored), we propose a method that (a) estimates the sample variance for the particular data frame, and then performs the subtraction from the Gram matrix.

An interpolation method is described for estimation of the frequency, amplitude and phase of a sinusoidal signal from the DFT of its discrete samples. It is shown that the scalloping loss, typically 1 db or more in tapered-window estimation, can be eliminated or substantially reduced. Also, robustness of this approach to noise is demonstrated by simulation examples and theoretical bounds. For multifrequency signals two different cases are treated. The first, where the frequencies of the sinusoids are known (such as in multifrequency radar detection), lends itself to closed form solution. In the second case, the constituent tone frequencies are estimated through interpolation, and then the amplitudes and phases are calculated as in the first case.

Complex-valued time series occur in many applications [1],[2] and frequently it is necessary to obtain a rational model for its autocorrelation function. A computationally efficient, noniterative method for developing such models is presented in this paper. The autocorrelation data samples are processed by a set of complex measurement units (first order recursive filters) to generate a family of auxiliary sequences. The Gram matrix of these sequences and the known impulse response of the measurement units is then shown to yield a parameter vector consisting of the coefficients of an ARMA model. Although similar to [3],[4] the solution is slightly different; also it permits analysis of complex signals. Results of application to SSB coherent radar returns are presented.

The ill-posed nature of direct deconvolution has prevented its use in applications such as channel equalization or inversion of transducer distortion. In this paper we discuss the use of a recursive nonlinear filter. The design of the proposed filter is accomplished in two stages. First, a constrained least-squares technique is employed to obtain, off-line, a filtered signal from a typical distorted signal. Then this pair of signals is used in a model identification scheme, to arrive at the nonlinear recursive filter. Since the final filter is recursive, and synthesized in one of the two alternative structures, both the memory and computation requirements are low enough to permit efficient on-line implementation.