The TMS320C8x family of Digital Signal Processors (DSP) incorporate multiple, high performance, fully programmable, processors onto a single CMOS die. The 320C80 has four 16/32-bit integer DSP CPUs, a 32-bit Floating Points RISC processor, 50 K bytes of SRAM, and a crossbar bus that can support over 4 Gigabytes of bandwidth. The 320C82 includes two of the integer DSPs, the RISC processor, 44 K bytes of SRAM, and a crossbar bus that can sustain over 2.6 Gigabytes of bandwidth per second. The C8x devices have special features that support general signal processing, telecommunication processing, and image processing. In addition to integrating multiple processors, the architecture includes on-chip Data RAM and instruction caches, as well as a very intelligent DMA controller, that support very high performance applications using lower cost off-chip memory. This paper concludes with a brief discussion of some of the future directions for this family of processors.

In this paper we examine the use of a recent innovation, called the logarithmic residue number system or LRNS, as an alternative to conventional DSP processors for implementing multiply-accumulate operations. We have fabricated a custom VLSI processor based upon this technology that is capable of providing substantial acceleration of vector arithmetic operations, convolution, correlation, and Fourier transforms in a relatively small, fast processor core when compared with implementations based on conventional arithmetic. The constituent arithmetic elements can be used as standard cells to implement application specific DSP designs.

This paper discusses the architectures of the TM-66 swiFFT DSP chip, the algorithms that the swiFFT chip is optimized for, the design of the S2 vector processing node that uses it, and applications of the above. The architecture of the swiFFT chip is optimized for real-time DSP processing in general, and the FFT algorithm in particular. It uses twenty IEEE 32-bit floating point adders and multipliers to sustain 680 MFLOPS during FFT processing and other algorithms, making it one of the fastest floating point DSP chips generally available. This paper discusses in detail the internal architecture, the tradeoffs between performance and generality, and the implementation of the FFT and other algorithms using this chip. In conclusion, some brief design overviews of a dual TM-66 board and an FFT building-block module are given to illustrate data flow through, and control of, the TM-66 swiFFT chip.

The hardware portion of the Harris Semiconductor Personal Computer Multimedia System is a 5 chip set which implements, in conjunction with a host processor and associated software and firmware, a complete H.320 video teleconferencing capability over ISDN 2B lines. The chip is comprised of a PAL/NTSC Video Encoder, a PAL/NTSC Video Decoder, a Video Codec, a Bus Interface and Audio Processor chip, and an Audio Codec. All 5 chips in the set are implemented in a 0.5 or 0.6 micron CMOS process. Each of the chips implement digital signal processing algorithms of varying levels of complexity and flexibility. These levels range from standard interpolation and decimation filter implementations found on the Audio Codec to dual programmable digital signal processor cores found on the Bus Interface and Audio Processor chip. A top level description of the chip set architecture is presented, along with a functional description of a typical video teleconferencing system based on this chip set. This is followed by a top level description of the various digital signal processing architectures and approaches used in the individual chips in the chip set.

Various video compression techniques are introduced. The basic idea of Discrete Cosine Transform and Motion Estimation compression is outlined, and a detailed example of the decoding of an H.261 video bitstream is demonstrated. More advanced techniques for DCT/ME compression as used in MPEG-2 and H.263 are described. A discussion of video encoders and decoders highlights the fact that `standard- compliant' video systems can yield vastly different results depending on implementation techniques, such as the algorithms used, the use of accelerating hardware.

With the success of V.34 wireline technology running at 28.8 Kbps over the PSTN (public switched telephone network), the next level of readily available high data rate communications to the home over twisted pair is ISDN (Integrated Services Digital Network) at 160 Kbps. But already the communication providers are anticipating the implementation of a new generation of high speed communications that will provide 6 Mbps to the home. And on the drawing boards of the ANSI T1E1.4 committee are plane for providing 54 Mbps to the home over twisted pair in conjunction with fiber to the curb technology. This paper discusses some of the issues associated with high data rate communications over twisted pair.

A digital filter which has been designed to be limit cycle free may exhibit limit cycles at the implementation stage. This is due to the inability to implement filter coefficients exactly in hardware when they are quantized to satisfy available wordlength requirements. Given a digital filter which is limit cycle free under zero input conditions, the work below presents an algorithm which finds a region in the coefficient space, about the nominal filter coefficient values, where in the filter remains limit cycle free. Furthermore the results of the algorithm will also indicate the availability of other machine representable numbers for the coefficients that fall within this robustness region. Hence one may even choose shorter wordlength registers for coefficient storage if the corresponding grid falls within the constructed robustness region.

High resolution direction finding techniques, such as MUSIC, generally do not perform well in the presence of correlated signals. In some situations, the direction of one (or more) of the correlated signals is known (such as when a `friendly' known signal is being jammed by an unknown signal). In this case, constraints can be included to block the direction of the known signal(s). Since the known direction usually have some uncertainty, additional constraints are necessary in order to block the known signals and hence decorrelate the unknown signal. This paper examines the use of derivation constraints to achieve signal decorrelation when the directions of `friendly' signals are known only approximately. This same concept of derivative constraints is also applied to the wideband signal case after using spatial resampling to focusing the correlation matrix. A cubic spline based resampling technique is proposed rather than the conventional sinc function resampling in order to achieve adequate resolution in the uncorrelated case before applying derivative constraints for correlated signals.

This paper is concerned with the problem of blind separation of independent signals (sources) from their linear convolutive mixtures. The problem consists of recovering the sources up to shaping filters from the observations of multiple-input multiple-output (MIMO) system output. The various signals are assumed to be linear but not necessarily i.i.d. (independent and identically distributed). The problem is cast into the framework of spatio-temporal equalization and estimation of the matrix impulse response function of MIMO channels (systems). An iterative, Godard cost based approach is considered for spatio-temporal equalization and MIMO impulse response estimation. Stationary points of the cost function are investigated and it is shown that all stable local minima correspond to desirable minima when doubly infinite equalizers are used. Analysis is also provided for the case when finite-length equalizers exist. The various input sequences are extracted and cancelled one-by-one. The matrix impulse response is then obtained by cross-correlating the extracted inputs with the observed outputs. Identifiability conditions are analyzed. Computer simulation examples are presented to illustrate the proposed approach.

For mobile communication, one essential component in base station design is the development of an adaptive array system to track low SNR signals buried in co-channel interferences. In this research, we examine three types of adaptive beamformers using the linear mean square, Kalman filter and minimum variance distortionless response design criteria to protect the desired signal and reject interferences. The purpose of this study is two folds. First, the performance of these algorithms are compared in terms of accuracy and speed of convergence. Second, we present a new tracking method based on beamforming to determine, on a continuous time basis, the direction of arrival of the desired signal. This approach will enable locating the source of the transmitted signals, nullifying the co-channel interferers and, hence, enhancing signal reception.

We introduce fundamentals of wavelet signal processing. We review the wavelet transform and multiresolution analysis. We illustrate their properties by simple examples.

Scalar-valued Malvar wavelets have been used to eliminate the blocking effects in scalar transform coding. In this paper, we introduce vector-valued Malvar wavelets for vector-valued signals. While constructing window vectors, we present a connection between vector-valued Malvar wavelets and vector Lemarie-Meyer band-limited wavelets. Similar to scalar-valued Malvar wavelets, vector-valued Malvar wavelets have applications in eliminating the blocking effects in vector transform coding.

A broad array of applications in the Public Switched Telephone Network (PSTN) require detailed information about type of call being carried. This information can be used to enhance service, diagnose transmission impairments, and increase available call capacity. The increase in data rates of modems and the increased usage of speech compression in the PSTN has rendered existing detection algorithms obsolete. Wavelets, specifically the Discrete Wavelet Transform (DWT), are a relatively new analysis tool in Digital Signal Processing. The DWT has been applied to signal processing problems ranging from speech compression to astrophysics. In this paper, we present a wavelet-based method of categorizing telephony traffic by call type. Calls are categorized as Voice or Data. Data calls, primarily modem and fax transmissions, are further divided by the International Telecommunications Union-Telephony (ITU-T), formerly CCITT, V-series designations (V.22bis, V.32, V.32bis, and V.34).

The change-point detection problem is one of the central problems in statistical inference and nonstationary signal processing. In this research, we incorporate the wavelet transform technique into the change-point detection framework and address several arising issues. We first apply the change detection algorithm in the wavelet domain, and discuss the advantages and disadvantages of the approach. Then, we consider the effect of down-sampling and the use of non-wavelet filters. Finally, we propose a new scheme for change detection based on the local energy feature, which shows some clear advantage of the wavelet transform approach.

This paper presents a wavelet based image coding method achieving high levels of compression. A multi-resolution subband decomposition system is constructed using Quadrature Mirror Filters. Symmetric extension and windowing of the multi-scaled subbands are incorporated to minimize the boundary effects. Next, the Embedded Zerotree Wavelet coding algorithm is used for data compression method. Elimination of the isolated zero symbol, for certain subbands, leads to an improved EZW algorithm. Further compression is obtained with an adaptive arithmetic coder. We achieve a PSNR of 26.91 dB at a bit rate of 0.018, 35.59 dB at a bit rate of 0.149, and 43.05 dB at 0.892 bits/pixel for the aerospace image, Refuel.

The paper presents an approach to the design of half-band discrete-time wavelets. This is accomplished through the use of a class of quadrature mirror filters which exhibit near- perfect reconstruction property. In particular, we present a technique for the design of such filters, wherein the designer has the flexibility to make tradeoffs between in- band behavior, out-of-band behavior, and the transition-band behavior. The basic formulation is carried out in the frequency domain, which is shown to translate the design problem into an eigenvalue-eigenvector problem. To find the optimal filter for a specific set of specifications, an optimization algorithm is also presented. Using this algorithm, designs ranging from 4 to 80 taps have been carried out successfully. A fairly complete table of resulting filters, which can be used by signal and image processing engineers, is included in the paper.

Recognition of 3D objects independent of size, position, and rotation is an important and difficult subject in computer vision. A 3D feature extraction method referred to as the Open Ball Operator (OBO) is proposed as an approach to solving the 3D object recognition problem. The OBO feature extraction method has the three characteristics of invariance to rotation, scaling, and translation invariance. Additionally, the OBO is capable of distinguishing between convexities and concavities in the surface of 3D object. The OBO also exhibits a good robustness to noise and uncertainty caused by inaccuracies in 3D measurements. A wavelet de- noising method is used for filtering out noise contained in the feature vectors of 3D objects.

This paper will describe the vector quantization (VQ) of speech spectrum parameters as used in state of the art real time speech coding algorithms. The complexity of these speech coding algorithms, combined with the constraints of real time operation, limit the practical size of the vector quantization tables. This conflicts with the requirement that large VQ tables are needed to provide high quality coded speech. The goal of this paper is to demonstrate the feasibility of using tree searched vector quantizers to overcome this conflict.

Mandarin Chinese is the official language in China and Taiwan, it is the native language of a quarter of the world population. As the services enabled by speech recognition technology (e.g. telephone voice dialing, information query) become more popular in English, we would like to extend this capability to other languages. Mandarin is one of the major languages under research in our laboratory. This paper describes how we extend our work in English speech recognition into Mandarin. We will described the corpus: Voice Across Taiwan, the training of a complete set of Mandarin syllable models, preliminary performance results and error analysis. A fast prototyping system was built, where a user can write any context free grammar with no restriction of vocabulary, then the grammar can be compiled into recognition models. It enables user to quickly test the performance of a new vocabulary.