To improve speech recognition performance, feature transformation based on discriminant analysis has been widely used to reduce the redundant dimensions of acoustic features. Linear discriminant analysis (LDA) and heteroscedastic discriminant analysis (HDA) are often used for this purpose, and a generalization method for LDA and HDA, called power LDA (PLDA), has been proposed. However, these methods may result in an unexpected dimensionality reduction for multimodal data. It is important to preserve the local structure of the data when reducing the dimensionality of multimodal data. In this paper we introduce two methods, locality-preserving HDA and locality-preserving PLDA, to reduce dimensionality of multimodal data appropriately. We also propose an approximate calculation scheme to calculate sub-optimal projections rapidly. Experimental results show that the locality-preserving methods yield better performance than the traditional ones in speech recognition.

Many signal processing systems, particularly in the multimedia and telecommunication domains, are synthesized to execute data-intensive applications: their cost related aspects -- namely power consumption and chip area -- are heavily influenced, if not dominated, by the data access and storage aspects. This paper presents an energy-aware memory allocation methodology. Starting from the high-level behavioral specification of a given application, this framework performs the assignment of the multidimensional signals to the memory layers -- the on-chip scratch-pad memory and the off-chip main memory -- the goal being the reduction of the dynamic energy consumption in the memory subsystem. Based on the assignment results, the framework subsequently performs the mapping of signals into both memory layers such that the overall amount of data storage be reduced. This software system yields a complete allocation solution: the exact storage amount on each memory layer, the mapping functions that determine the exact locations for any array element (scalar signal) in the specification, and an estimation of the dynamic energy consumption in the memory subsystem.

To precisely model the time dependency of features is one of the important issues for speech recognition. Segmental unit input HMM with a dimensionality reduction method has been widely used to address this issue. Linear discriminant analysis (LDA) and heteroscedastic extensions, e.g., heteroscedastic linear discriminant analysis (HLDA) or heteroscedastic discriminant analysis (HDA), are popular approaches to reduce dimensionality. However, it is difficult to find one particular criterion suitable for any kind of data set in carrying out dimensionality reduction while preserving discriminative information. In this paper, we propose a new framework which we call power linear discriminant analysis (PLDA). PLDA can be used to describe various criteria including LDA, HLDA, and HDA with one control parameter. In addition, we provide an efficient selection method using a control parameter without training HMMs nor testing recognition performance on a development data set. Experimental results show that the PLDA is more effective than conventional methods for various data sets.

The checkerboard effect is caused by the periodic time-variant property of multirate filters which consist of up-samplers and digital filters. Although the conditions for some one-dimensional (1D) multirate systems to avoid the checkerboard effect have been shown, the conditions for Multidimensional (MD) multirate systems have not been considered. In this paper, some theorems about the conditions for MD multirate filters without checkerboard effect are derived. In addition, we also consider MD multirate filter banks without checkerboard effect. Simulation examples show that the checkerboard effect can be avoided by using the proposed conditions.

This paper proposes a three-dimensional (3-D) band-limiting technique for a conversion of Simple Cubic Sampling into Body-Centered Cubic Sampling. Based on spectral distribution of the original signal, the proposed method adaptively varies the passband shape of a band-limiting filter in order to preserve informations of the original picture. By applying this method to 3-D moving pictures, we can preserve resolution on each axis without introducing heavy aliasing error and avoid degradation of picture quality such as ringing in still pictures or blurring in moving pictures. The examples given in this paper demonstrate these advantages.

This paper proposes a design technique for 3-D non-separable QMF banks with Face-Centered Cubic Sampling (FCCS) and Body-Centered Cubic Sampling (BCCS). In the proposed technique, 2-D McClellan transformation is applied to a suitably designed 2-D prototype QMF to obtain 3-D QMFs. The design examples given in this paper demonstrate advantages of the proposed method.

In this paper, we propose two parallel processing methods for multidimensional (MD) sampling lattice alteration. The use of our proposed methods enables us to alter the sampling lattice of a given MD signal sequence in parallel without any redundancy caused by up- and down-sampling, even if the alteration is rational and non-separable. Our proposed methods are provided by extending two conventional block processing techniques for FIR filtering: the overlap-add method and the overlap-save method. In these proposed methods, firstly a given signal sequence is segmented into some blocks, secondly sampling lattice alteration is implemented for each block data individually, and finally the results are fitted together to obtain the output sequence which is identical to the sequence obtained from the direct sampling lattice alteration. Besides, we provide their efficient implementation: the DFT-domain approach, and give some comments on the computational complexity in order to show the effectiveness of our proposed methods.

This paper discusses a new design method for 2-D variable FIR digital filters, which is an extension of our previous work for 1-D case. The method uses a 3-D prototype FIR filter whose cross-sections correspond to the desired characteristics of 2-D variable FIR filters. A 2-D variable-angle FIR fan filter is given as a design example.

This paper presents an optimal filtering algorithm using the covariance information in linear continuous distributed parameter systems. It is assumed that the signal is observed with additive white Gaussian noise. The autocovariance function of the signal, the variance of white Gaussian noise, the observed value and the observation matrix are used in the filtering algorithm. Then, the current filter has an advantage that it can be applied to the case where a partial differential equation, which generates the signal process, is unknown.

In a previous article, an optical array imaging system has been presented. In this system, first, a set of array data is collected by repeatedly illuminating the object with laser light from each array element, detecting the reflected light as interferogram, and extracting the reflected wave field based on the spatial heterodyne detection. Then, an eigenvalue analysis is applied to the data to derive the wave field that would backpropagate and focus at a single point on the object; in this case, the iterative algorithm is used which indicates that the object point may have the largest reflectivity. It was shown experimentally that the single-point-focusing was attained for objects having several such parts with almost the same reflectivities. A preliminary study by computer simulation, however, indicates that the probability with which the wave focuses at multiple object points would not be small enough, resulting in a degraded image with ghost image components. In this paper, the array data within subaperture regions are selectively used to attain the single-point-focusing and obtain a good image for any object. First, it is shown analytically that the change in the dimension or center position of the aperture is effective to change the eigenvector so that it attains the single-point-focusing. Then, a procedure to find the optimum subapertures and a measure evaluating the degree of single-point-focusing for the eigenvector are presented. The method is examined in detail using experimentally obtained array data, and the results show that the method is effective in obtaining good images for any objects without sacrificing image resolution. When we compare the imaging system to an automatic focusing camera, it may be said that the additional processings enhance the capability of automatic focusing to a great degree.

In this paper, a new description of a separable-denominator (S-D) two-dimensional (2-D) transfer matrix is proposed, and its realization is considered. Some of this problem had been considered for the transfer matrices whose elements are two-variables rational functions. We shall propose a 2-D transfer matrix whose inputs-outputs relation is represented by a ratio of two-variables polynomial matrices, and present an algorithm to obtain a 2-D state-space model from it. Next, it is shown that the description proposed in this paper is always minimally realizable. And, we shall present a method of obtaining the description proposed in this paper from a S-D 2-D rational transfer matrix.

An optical array imaging system is presented with basic experimental results. First, a remote object is illuminated with laser light at an angle and the reflected light is detected with an array sensor after interfering it with the reference light. This process is repeated by changing the illumination angle to collect a set of fringe patterns, which are A/D converted and stored in a harddisk in a computer. Then, the data are processed on a computer, first, to estimate the complex-amplitude object wave fields, second, to derive the eigenvector with the maximum eigenvalue for the correlation of the estimated object fields, and finally, to form an image of the object. The derivation of the eigenvector follows an iterative algorithm, which can be interpreted as the process of repeating backward wave propagation of the field between the two apertures illuminating and detecting laser light. The eigenvector field can be expected to backpropagate to focus at a point on the object with the maximum coefficient of reflection, so that a beam-steering operation is applied to the eigenvector to form an image of the object. The method uses only the information of the array data and the lateral spacings of the receiving array (CCD) elements. Hence, the method can give good images of objects even if the reference light is uncollimated with an unknown distorted wavefront, and even if the illuminating angles are imprecise in three dimensions. Basic experimental results clearly show the usefulness of the method.