Attention is increasing on the need to educate students in areas that extend beyond the traditional technical basis for their field: the Kellogg Commission's mandate for a "scholarship of engagement" [3]; the Engineering Deans Council report's emphasis on teamwork, communication, and the need for students to "understand the economic, social, environmental and international context of their professional activities" [2]; the CSAB Criteria 2000 requirements in written and oral communication and "coverage of social and ethical implications of computing" [1]; the growth of broad IT programs and schools.Service learning offers the opportunity to broaden the educational experience by engaging students in "real-world" projects. However, it also poses new questions: Should students earn academic credit for courses in which a significant portion of their time is spent on "soft" skills? Are students gaining valid technical experience on projects defined by an outside customer rather than by CS&E faculty? How can curriculum structures support projects that aren't constructed to fit within semester boundaries?The panel includes faculty and students from service-learning programs, the chief technical officer for a software company, and a community "customer."

We describe a prototype spoken language system that loosely integrates a speech recognition component based on hidden Markov models with a constraint dependency grammar (CDG) parser using a word graph to pass sentence candidates between the two modules. This loosely coupled system was able to improve the sentence selection accuracy and concept accuracy over the level achieved by the acoustic module with a stochastic grammar. Timing profiles suggest that a tighter coupling of the modules could reduce parsing times of the system, as could the development of better acoustic models and tighter parsing constraints for conjunctions.

One of the most challenging problems in engineering-and signal processing-education is providing realistic and meaningful design experience. In the Engineering Projects in Community Service (EPICS) program, teams of engineering undergraduates earn academic credit for multi-year projects that solve technology-based problems for community organizations. Key features of EPICS include the long-term nature of the projects; emphasis on "real-world" start-to-finish design; the learning experience embodied in solving ambitious engineering problems; vertical, multidisciplinary teams; development of teamwork and communication skills; and the use of engineering to help the community. We describe the EPICS program and highlight four EPICS signal processing projects: a real-time system to measure speaking rate for Purdue's speech clinic; voice-controlled interactive software to encourage speech in developmentally delayed children; a microphone array hearing aid; and a virtual museum tour and interactive Web-based history games for the Tippecanoe County Historical Association.

The research presented here focuses on implementation and efficiency issues associated with the use of word graphs for interfacing acoustic speech recognition systems with natural language processing systems. The effectiveness of various pruning methods for graph construction is examined, as well as techniques for word graph compression. In addition, the word graph representation is compared to another predominant interface method, the N-best sentence list. 1. INTRODUCTION An important research topic in recent years has been the integration of speech recognition systems with language models [2, 8]. Many systems integrate stochastic language models directly into the speech recognizer. However, a structure in which a front-end acoustic recognizer is interfaced to a separate language processing module allows use of more sophisticated parsing techniques and additional semantic and contextual information to aid in speech understanding. The choice of data representations used to accomplish ...

We present analytical and experimental results for fine-grained list ranking algorithms. We compare the scalability of two representative algorithms on random lists, then address the question of how the locality properties of image edge lists can be used to improve the performance of this highly data-dependent operation. Starting with Wyllie's algorithm and Anderson and Miller's randomized algorithm as bases, we use the spatial locality of edge links to derive scalable algorithms designed to exploit the characteristics of image edges. Tested on actual and synthetic edge data, this approach achieves significant speedup on the MasPar MP-1 and MP-2, compared to the standard list ranking algorithms. The modified algorithms exhibit good scalability and are robust across a wide variety of image types. We also show that load balancing on fine grained machines performs well only for large problem to machine size ratios.

This paper presents a new statistical speech model in which coarticulation is modeled explicitly. Unlike HMMs, in which the current state depends only on the previous state and the current observation, the proposed model supports dependence on the previous and next states and on the previous and current observations. The degree of coarticulation between adjacent phones is modeled parametrically, and can be adjusted according to a parameter representing the speaking rate. The model also incorporates a parameter that represents a frame-by-frame measure of confidence in the speech. We present two methods for solving the system parameters: one based on the K-means method, and a novel method based on explicitly minimizing a measure of the segmentation error. A new, efficient forward algorithm and the use of top candidates in the search greatly reduce the computational complexity. In evaluation on the TIMIT data base, we achieve a phone recognition rate of 77.1%.

Hybrids that use a neural network to estimate the output probability for a hidden Markov model (HMM) word recognizer have been competitive with traditional HMM recognizers when both use monophone context. While traditional HMM recognizers can easily utilize more context (e.g., triphones) to achieve better results, the size of the task has made it impractical to use phonetic context directly in the neural network front end of a hybrid. In this paper, we suggest a simple method to incorporate more context by modeling the phone distributions obtained from the neural network. This allows the HMM to easily handle stochastic pronunciations as well as errors from the neural network phone recognizer. The re-estimation equations are derived for the new model. Results for the Resource Management task illustrate that SOHMM increases recognition accuracy for the cases of no grammar, unigram grammar, and word pair grammar.

We present a study on the use of lexical stress classication to aid in the recognition of phonetically similar words.In this study, w e use a simple pattern recognition approach t o determine which syllable is lexically stressed for phonetically similar word pairs (e.g., PERfect, perFECT) extracted from continuously spoken sentences.We use a combination of two features from the acoustic correlates of lexical stress, and assume multivariate Gaussian distributions to form a Bayesian classier.The features used are normalized energy and duration of the vowel for each syllable of the word.We e v aluate several normalization methods.Two sets of sentences were designed for this study.F or the pilot experiment, the classication accuracy on words from the natural sentence set was 89.9% and on words from the control sentence set was 100%.To improve the performance, three-feature classiers, which included two normalized energy features and one normalized duration feature, were developed.The classication accuracy on words from the natural sentence set was 97.23%.

We present a probabilistic error correction technique to be used with an average magnitude dierence function (AMDF) based pitch detector.This error correction routine provides a v ery simple method to correct errors in pitch period estimation.Used in conjunction with the computationally ecient AMDF, the result is a fast and accurate pitch detector.In performance tests on the CSTR (Center for Speech Technology Research) database, probabilistic error correction reduced the gross error rate from 6.07% to 3.29%.

We show that many of the errors in a context-dependent phone recognition system are due to poor segmentation. We then suggest a method to incorporate explicit segmentation information directly into the HMM paradigm. The utility of explicit segmentation information is illustrated with experiments involving five types of segmentation information and three methods of smoothing.

Introduces a new recursion that reduces the complexity of training a semi-Markov model with continuous output distributions. It is shown that the cost of training is proportional to M/sup 2/+D, compared to M/sup 2/D with the standard recursion, where M is the observation vector length and D is the maximum allowed duration.>

We describe a project that integrates applications requirements, parallel algorithm design, models of parallel computing, and software tools in order to improve the ability of applications researchers in the fields of computer vision and image processing (CVIP) to realize the performance potential of high performance parallel computers. This objective is achieved by pursuing four directions of research: the development of efficient and practical scalable algorithms for fundamental CVIP problems; the implementation of realistic CVIP scenarios on high performance computers; the development of a scalable, architecture-independent parallel model and a metric for estimating the performance of an algorithm on an existing parallel machine; and the development of prototype software tools. In this paper we describe work in each of these areas.

In this paper, we present efficient parallel solutions for structural indexing-based object recognition on coarse-grained parallel machines. Based on the analysis using the C/sup 6/-model, the parallel algorithm proposed in this paper takes O(Sk/p) computation units and O(p/sup 3/2/) communication units on a p processor coarse-grained machine such that 1/spl les/p/spl les/S, whereas the sequential solution takes O(Sk). The proposed solution is implemented on the Intel Touchstone Delta and performance results are shown. For a scene consisting of 160 feature points, the recognition phase takes 258 ms on a 64 processor Delta when the model database contains 256 models. The sequential algorithm takes approximately 4 seconds on a single node of the Delta.

We show that HMM word recognition using Deller and Snider's ( see ibid., vol.1, no.4, 1993) "any path" procedure makes an assumption of independence that is not made by either the forward or Viterbi algorithms. We also point out that additional savings in execution time can be achieved by precomputation.>

We have extended Maruyama's [5, 6, 7] constraint dependency grammar (CDG) to process a lattice or graph of sentence hypotheses instead of separate text strings. A post-processor to a speech recognizer producing N-best hypotheses generates the word graph representation, which is then augmented with information required for parsing. We will summarize the CDG parsing algorithm and then describe how the algorithm is extended to process a word graph on a single processor machine.

A stochastic model of English speech is described that can be used to generate a stream of phonemically labeled frames with the same characteristics as real speech to be used during the design of new computer architectures. Specialized computer architectures to perform speech understanding in real time should incorporate information about the characteristics of English speech. The English speech model described can be used to simulate efficiently the performance of new architectures and eliminates the need for massive amounts of input speech data to determine architecture design parameters. A stochastic model is used to reproduce accurately the phoneme content of English speech, including the transition probabilities, the duration mean, and the duration variance of each phoneme and silence region.>

A system for dynamic intelligent scheduling and control (DISC) of reconfigurable parallel processors is presented. The purpose of the system is to provide a rapid prototyping capability for computer vision/image processing tasks. The scheduler particularly addresses the problems of algorithms with execution times that depend on the image data and processing scenarios that vary dynamically based on the input image. Since conventional scheduling methods cannot propose schedules for most masks of this type, a dynamic controller is used to schedule the task and reconfigure the machine on the fly. This dynamic scheduling system attempts to balance the overall processing scenario with the needs of the individual routines that make up the task. The implementation of this system is discussed, with emphasis on the scheduling heuristics and the use of the system for prototyping computer vision/image processing tasks. Testing was done on a number of tasks that exercised different aspects of the scheduling strategy.>

An experimental analysis of the architecture of an SIMD/MIMD parallel processing system is presented. Detailed implementations of parallel fast Fourier transform (FFT) programs were used to examine the performance of the prototype of the PASM (Partitionable SIMD/MIMD) parallel processing system. Detailed execution-time measurements using specialized timing hardware were made for the complete FFT and for components of SIMD, MIMD, and barrier-synchronized MIMD implementations. The component measurements isolated the effects of floating-point arithmetic operations, interconnection network transfer operations, and program control overhead. The measurements allow an accurate extrapolation of the execution time, speedup, and efficiency of the MIMD, SIMD, and barrier-synchronized MIMD programs to a full 1024-processor PASM system. This constitutes one of the first results of this kind, in which controlled experiments on fixed hardware were used to make comparisons of these fundamental modes of computing. Overall, the experimental results demonstrate the value of mixed-mode SIMD/MIMD computing and its suitability for computational intensive algorithms such as the FET.>

A detailed study is presented of parallel fast Fourier transform programs executing on the prototype of the PASM parallel processing system. Detailed execution time measurements using specialized timing hardware were made for complete fast Fourier transforms (FFTs) and for SIMD, MIMD, and hybrid SIMD/MIMD implementations. The component measurements isolated the effects of floating-point operations, interconnection network transfer operations, and program control overhead. Using these detailed measurements, an expression for projecting the execution time for an M-point FFT executing on M/2 PASM processing elements is derived. This expression is then used to obtain an accurate extrapolation of the execution time and speedup of the MIMD, SIMD, and hybrid programs to a full 1024-processor PASM system. Overall, the experimental results demonstrate the value of the multimode capability of PASM and the stability of PASM for computationally intensive algorithms such as the FFT.>

The major goal of this research is to reduce the discrepancy in recognition performance between normal and abnormal speech, given that reference templates were derived only from normal speech. A method is devised that uses the differences in spectral slope between linear predictive coding log magnitude spectra to weight the point-by-point energy differences between the spectra. The distances of all reference tokens of like phonemes are combined to form a smallest cumulative distance (SCD) method. When SCD is combined with the method of slope-dependent weighting (SDW), the most significant success is obtained. In terms of error rates for a fixed phoneme vector length of five, SDW+SCD is found to reduce the difference in error rate between normal and abnormal speech by approximately 50%.>

A parallel algorithm for extracting the roots of a polynomial is presented. The algorithm is based on Graeffe's method, which is rarely used in serial implementations, because it is slower than many common serial algorithms, but is particularly well suited to parallel implementation. Graeffe's method is an iterative technique, and parallelism is used to reduce the execution time per iteration. A high degree of parallelism is possible, and only simple interprocessor communication is required. For a degree-n polynomial executed on an (n+1)-processor SIMD machine, each iteration in the parallel algorithm has arithmetic complexity of approximately 2n and a communications overhead n. In general, arithmetic speedup is on the order of p/2 for a p-processor implementation.>

The use of two parallel architectures, a single-instruction-stream, multiple-data-stream (SIMD) machine and a VLSI processor array, to implement an isolated word recognition system is examined. SIMD and VLSI processor array algorithms are written for each of the components of the recognition system. The component parallel algorithms are simulated along with two complete recognition systems, one composed of SIMD algorithms and other composed of VLSI processor array algorithms. The simulations show that a SIMD machine with 100 processing elements, each based on an 8-MHz MC68000, can perform isolated word recognition over a large vocabulary in real time. The VLSI processor array simulations show that an array using 51 cells, each containing a 12-MHz Intel 8051, can recognize speech in real time over a small vocabulary. The simulation revealed certain architectural features that can enhance the overall performance of the speech system.>

Speaker and phoneme recognition is investigated based on the discrete rectangular wave transform (DRWT). The DRWT is one of the discrete Fourier preprocessing transforms (DFPTs) which corresponds to the first stage in two-stage algorithms to represent discrete trigonometric transforms such as the discrete Fourier transform (DFT). The DRWT is much faster and easier to implement than the DFT. The recognition experiments performed also show that performance of recognition based on DRWT speech spectrograms is comparable to the performance of recognition based on DFT speech spectrograms.>

Acoustic-phonetic differences between normal and abnormal speech in reference to recognition systems for the aircraft cockpit are studied. The speech used in this research was collected in an anechoic chamber with an oxygen mask, and the speaker spoke (1) normally, (2) loudly ( approximately 10 dB above normal), and (3) with 90 dB of pink noise injected into the ears (invoking the Lombard reflex). Eighteen features were analyzed for approximately 11000 phonemes (from five speakers) to determine significant differences in the three types of speech. The most reliable trends observed were energy migrations in the frequency domain and the associated changes in spectral tilt for the sonorants.>

Some experiments in stop consonant recognition using a new speech analysis technique are presented. The speech analysis technique used a Mellin-Fourier Homomorphism (MFH) on the Fourier transform of initial stop consonant bursts. A template-matching K-nearest neighbor algorithm, with MFH spectra as input, was used in identifying the stops. Extensive experiments using 32 speakers (8 each of male adults, female adults, male children, and female children) and 9 vowel environments were run. Recognition accuracies as high as 85% for voiced stops and 94% for unvoiced stops were achieved without any preliminary voicing decision.

This paper examines the use of dataflow parallel processing for the execution of digital filtering algorithms. The dataflow language used is Lucid. FIR and recursive filtering algorithms are formulated in Lucid. Analysis of the parallelism achievable in the dataflow implementations is presented.

As the complexity of speech recognition systems increases, conventional computers are unable to perform all the operations quickly enough to process the speech input in real time. This paper examines the use of a single instruction stream - multiple data stream (SIMD) parallel architecture to reduce the computation time for a dynamic time warping template matching based isolated word recognition system. Each of the components of the recognition system was written as an SMD parallel algorithm. The component SIMD algorithms and the complete word recognition system were simulated. The simulations showed that with 100 processing elements, each an 8 MHz MC68000, the speech system can perform isolated word recognition over a large vocabulary in real time. This real-time speech system used a 20 KHz sampling rate, 16 bits per sample, 100 samples per frame, 8 LPC coefficients per frame, 16 bits per LPC coefficient, and a 1,000 word vocabulary.