We apply the maximum mutual information (MMI) criterion to discriminative training of hidden Markov model (HMM) parameters. In contrast to the conventional MMI training approach, we adopt the cross-validatory strategy with which the parameters are estimated on a part and assessed on the other parts of the training data. For this purpose, we propose the deleted MMI training method, which performs cross-validatory parameter updating while maintaining the converging behavior of the conventional MMI-based algorithm. The proposed method is compared to the conventional MMI approach in classification of artificial data and in speaker-independent continuous speech recognition, and shows better performance.

Below 4.8 kbits/s, code-excited linear prediction (CELP) coders in general suffer from two kinds of perceptually important degradation. They are inter-harmonic noise and high frequency mismatch. To remedy these degradations, we propose in this paper an improved weighting function which utilizes the spectral weighting methodology and also takes into account the periodic character in voiced sound. The function can adapt to variation of pitch by itself without any pitch estimation in voiced sound and is also applicable to all speech segments without any voiced/unvoiced discrimination algorithm. Simulation results show that the performance of the CELP coder with the proposed weighting function is better than that of the conventional CELP coder.

In this paper, we consider the hidden Markov model(HMM) parameter compensation in noisy environments with multiple noise sources based on the vector Taylor series(VTS) approach. General formulations for multiple environmental variables are derived and systematic expectation-maximization(EM) solutions are presented in maximum likelihood(ML) sense. It is assumed that each noise source is independent and having Gaussian distribution. To evaluate proposed method, we conduct speaker independent isolated word recognition experiments in various noisy environments. Experimental results show that proposed algorithm ahieves significant improvement. Especially, the proposed method is consistently more effective than the parallel model combination(PMC) based on log-normal approximation.

We present a novel method to incorporate temporal correlations into a speech recognition system based on conventional hidden Markov models (HMMs). The temporal correlations are considered to be useful for recognition because of the fact that the speech features of the present frame are highly informative about the feature characteristics of neighboring frames. In this paper, by treating these correlations in the form of conditional probability distributions (PDs), we propose a new technique for incorporating frame correlations. With the proposed method called the extended logarithmic pool (ELP), we approximate a joint conditional PD by separate conditional PDs associated with respective conditions. We provide a constrained optimization algorithm with which we can find the optimal value for the pooling weights. For practical purposes, we also suggest methods to get robust PD estimates for characterizing frame correlation. In addition, to improve model discriminability, a technique to combine two kinds of PDs through the exponents is introduced. The results in the experiments of speaker-independent continuous speech recognition with the proposed approaches show error reduction up to 20.5% as compared to that with the conventional bigram-constrained (BC) HMM method.

The output probability distributions (PDs) in each state of a discrete HMM suffer from sparseness, causing inaccurate modeling of probabilistic characteristics of speech features within the state. A desirable solution to the problem arising from insufficient training data is to interpolate a maximum likelihood (ML) estimate of a PD with some other estimates that are, to some extent, able to strengthen the robustness of the PD with respect to unseen data. We propose a statistically reliable deleted interpolation (DI) approach. The DI is an efficient technique for interpolating several probability distribution (PD) estimates, and usually different interpolating weights are used for each predetermined range of PD counts. Our approach attempts to piecewise linearly approximate the interpolating weight curve based on some reasoning concerned with statistical reliability of sample-based estimates.

We propose a new method of finding discriminative state-weights recursively using the generalized probabilistic descent method. This method is implemented with minor modification of the conventional parameter estimation and recognition algorithms by constraining the sum of the state-weights to the number of states in a recognition unit, and can be applied to continuous speech recognition as well as isolated word recognition. We confirm the validity of the method with phoneme-based and word-based state-weighting schemes for three kinds of recognition tasks.

We propose a parallel-branch subunit model in semicontinuous HMM for improved initialization and training in word recognition. We obtain the model by adding a new subunit branch based on misrecognized data in training to the previous parallel branches for that subunit. Simulation results show that this proposed model is efficient and gives good recognition performance in word recognition.

In this paper, we propose two methods of obtaining a parallel-branch subunit model for improved recognition performance. In one method, the model is obtained by adding a new subunit branch based on misrecognized data in training to the previous parallel branches for that submit. In the other method, it is obtained by splitting off each subunit branch based on mixture components in continuous hidden Markov model. We propose to use a good initialization point obtained by error corrective estimation rather than by random or probabilistic statistics in the parallel-branch model when the number of mixtures and parallel branches increases.

We propose a modified hidden Markov model (MHMM) that incorporates nonparametric state duration and state duration-dependent observation probabilities to reflect state transitions and to have accurate temporal structures in the HMM. In addition, to cope with the problem that results from the use of insufficient amount of training data, we propose to use the modified continuous density hidden Markov model (MCDHMM) with a different number of mixtures for the probabilities of state duration-independent and state duration-dependent observation. We show that this proposed method yields improvement in recognition accuracy in comparison with the conventional CDHMM.

We present various methods for estimating a robust output probability distribution (PD) in speech recognition based on the discrete hidden Markov model (HMM). In speech recognition, we encounter the problem of an insufficient amount of training data, which may cause inaccurate modeling of the HMM parameters, especially the output PD's. In this paper, to enhance the robustness of the output PD's with respect to unseen data, we study two approaches: smoothing and tying of the PD's. We introduce a new algorithm to smooth a PD, where a smoothing matrix is estimated by following the strategy of cross-validation as used in deleted interpolation. As for tying, we derive a number of state classes based on a clustering tree which achieves a good compromise between robustness and detail of the tied PD's in specifying speech feature characteristics. In addition to providing an efficient method for constructing the clustering tree, we suggest a measure that accounts for the variation of estimated PD's under various situations. The performances of the proposed methods are evaluated by speaker-independent isolated word recognition experiments and are shown to be better in recognition accuracy than that of the PD's based on the maximum likelihood criterion.>

In a multiwavelength optical network with tunable transceivers, an algorithm to make the transmitter and the receiver tuned to the same wavelength simultaneously is needed. The paper proposes contention-based reservation protocols using a separate control channel for multiwavelength optical networks with a passive star topology. First, they present a protocol which can be used in the network where the ratio of the end-to-end signal propagation delay to the transmission time of a data packet is smaller than 1. Also, for a very high speed network where this ratio is greater than 1, the authors present three protocols according to the variability of the packet length and the buffering of the reservation. To access the control channel, all the proposed protocols use the slotted ALOHA protocol. The authors analyze these protocols with a finite population model and investigate the delay-throughput characteristics.

The asynchronous transfer mode (ATM) has been selected as the multiplexing and switching technique for use in the public broadband ISDN (B-ISDN). We propose a large-scale ATM switch architecture in which a banyan multipath self-routing network is combined advantageously with a shared buffer type switch element. The proposed banyan space-division concept yields a simple architecture having the potential to accommodate easily the growth of switch size. Since the interconnection network between switch modules or between switch elements has a twofold banyan architecture, expansion in crosspoints or interconnections with the increase of switch size can be lessened. The multipath self-routing concept makes the switch performance better and leads to an efficient realization of a switch element on a single chip as the fundamental building block of a large-size switch. We analyze the required capacity for queuing buffers in the switching network. The multipath approach inevitably creates information sequence disturbances. Therefore, we also analyze the out-of-sequence phenomenon of a banyan multipath switching system. To satisfy the sequence integrity requirement for ATM, a simple approach is proposed for the multipath switch by using a spacing controller. In addition, we quantify the improvement of out-of-sequence performance under the spacing controller scheme.

In this paper, we introduce a Korean large vocabulary speech recognition system. This system recognizes sentence utterances with a vocabulary size of 1160 words, and is designed for an automatic telephone number query service. The system consists of four subsystems. The first is an acoustic processor recognizing words in an input sentence by a Hidden Markov Model (HMM) based speech recognition algorithm. The second subsystem is a linguistic processor which estimates input sentences from the results of the acoustic processor and determines the following words using syntactic information. The third is a time reduction processor reducing recognition time by limiting the number of candidate words to be computed by the acoustic processor. The time reduction processor uses linguistic information and acoustic information contained in the input sentence. The last subsystem is a speaker adaptation processor which quickly adapts parameters of the speech recognition system to new speakers. This subsystem uses VQ adaptation and HMM parameter adaptation based on spectral mapping. We also present our recent work on improving the performance of the large vocabulary speech recognition system. These works focused on the enhancement of the acoustic processor and the time reduction processor for speaker-independent speech recognition. A new approach for speaker adaptation is also described.

Transmission errors are known to be bursty in nature. In block-oriented data communication, this bursty error behavior should be treated assuming that there are non-renewal correlations between the transmission conditions of data block. In this paper, using the stochastic sequential machine method the authors suggest a block-level burst error model of a digital tandem connection that consists of a set of segmented channels, and then investigate several end-to-end error performance characteristics under the burst-error condition. Also, they derive expressions that approximate the tandem connection model parameters by combining individual segment channel models into an equivalent combined error model. They show the proposed method is very useful in describing the error performance of tandem link connections.>

The authors investigate the effect of forward/backward channel memory (statistical dependence in the occurrence of transmission errors) on ARQ error controls. To take into account the effect of backward channel errors in the performance analysis, they suppose some modified ARQ schemes that have an effective retransmission strategy to prevent the deadlock incurred by the errors on acknowledgments. They consider two modified go-back-N schemes with timer control and with buffer control, one modified selective-repeat scheme and one modified stop-and-wait scheme. By analytical and simulation results, they show how much the performance in throughput and delay of the modified ARQ schemes is affected by the degree of error burstiness when the forward and backward channels are governed by a first-order Markovian block error pattern. They also show that there exists a large difference between the performances with the burst and independent error cases.>

This paper presents an exact analysis of a general asymmetric polling system which has a finite-capacity buffer and exhaustive service discipline at each station. To analyze the system we consider a model system in which each station is assumed to have a temporary virtual buffer with infinite capacity attached to the real buffer, while the server is unavailable. It is introduced to store all the messages which are to be lost due to the limitation of the real buffer capacity during the vacation period. At the instant when the server visits a station, using the marginal queue length distribution of the station including the virtual buffer, the intervisit time distribution of the station is derived, The system behavior using the relationship between performance measures and the intervisit time distribution is then investigated.>

Analyzes the performance of a statistical multiplexer for heterogeneous types of bursty traffic in an asynchronous transfer mode (ATM) network. The authors model each type of bursty traffic with a Markov-modulated Poisson process (MMPP), and represent the different burstinesses of traffic by changing the parameters of the MMPP. Then, they use a finite queue multi-server model for analysis of the statistical multiplexer with multiple output channels. In order to consider the slottized operation of the ATM channel, they also assume that transmission of the cells queued in the buffer occurs at discrete time instants. With these traffic and system models they obtain the cell delay distribution and the cell blocking probability for the superposed heterogeneous types of bursty traffic. They also investigate the performance of each traffic separately by obtaining the cell delay and the cell blocking probability. From numerical results, they show the effects of each traffic's characteristic on the performance of the statistical multiplexer. According to these results, it is noted that some efficiencies might be effected by dividing those types of traffic with comparable burstiness into segregated groups and allocate bandwidth to them exclusively.>

The authors present a model for the broadcast star network with collision avoidance switch and with multiple priority classes of messages. They also propose a service discipline which is easy to implement and exhibits relatively low variance of message delay. This service discipline, which is used to select a message among messages with the same class-index, is a combination of first-come first-served (FCFS) and random-order of service (ROS) disciplines, and it uses the priority fields in the message header. Then, the ready stations can be considered to form a distributed queue, and the entire system can be considered to form a priority M/G/1 queueing system with multiple server vacations. They analyze the model to obtain the distributions of the message delay for the proposed service discipline, and compare the results with that for the FCFS discipline. From numerical results, they show that the variance of the message delay for the proposed service discipline is almost the same as that for the FCFS discipline for any distribution of message length and also for any value of normalized propagation delay.>

Analyzes the performance of the unslotted 1-persistent carrier sense multiple access with collision detection (CSMA-CD) protocol for fiber optic bus networks with a finite number of stations M, each of which has an infinite storage buffer for storing packets with its length fixed or exponentially distributed. Treating the bus network as an M-class priority queueing system with a preemptive repeat-identical retransmission strategy, the authors show numerically that this 1-persistent retransmission strategy provides better delay-throughput performance than the non-persistent one. To satisfy the fairness of equal delays among stations in the bus network, the authors next present the unslotted p/sub i/-persistent CSMA-CD protocol. Under this protocol, the first packet arriving at station i is forced to enter into the backoff procedure with the probability of (1/spl minus/p/sub i/). After backoffs, transmission begins until the station becomes empty. The authors determine the persistent probability p/sub i/ and show that the delay-throughput performance of the p/sub i/-persistent strategy is better than that of the 1-persistent one, and further it is robust for any packet lengths. Simulation results are presented to verify the analytical results.>

Three call handling schemes for a base station which can handle both originating calls (OCs) and prioritized handoff calls (HCs), and which serves these calls under the blocked-call-delayed basis, are presented and compared. Delayed calls are assumed to be stored in a finite storage buffer. One scheme (Scheme I) exclusively allows handoff calls to be stored and to use a fixed number of guard channels. The other two schemes without guard channels (Scheme II-using the last-in/first-out policy for OC's and Scheme III-using the first-in first-out policy for OC's) allow both types of calls to be stored, and allow prioritized handoff calls to pushout originating calls if the buffer is full. For these three schemes, the blocking probabilities and delay distributions of both types of calls are numerically obtained, and a boundary for call handling schemes between with and without guard channels is found. From the numerical results, it is shown that Scheme II can be considered a good candidate for a call handling scheme for microcellular systems.>

The authors propose a combined preemptive/nonpreemptive priority discipline. When a high-priority job arrives at the system while a low-priority job is being in service, the high-priority job will be served immediately and the low-priority job will go back to the head of the queue of its class, if a discretion rule for preemption is satisfied. Otherwise, the high-priority job waits in queue until the completion of the low-priority job service. As the discretion rule for preemption, three schemes are considered, each based on the parameter of the low-priority job: the elapsed service time, the ratio of elapsed to total service time, and the remaining service time. Using the busy-period analysis technique, an M/G/1 queueing system with multiple priority classes of jobs is analyzed. Considered preemptive rules are the preemptive-resume and preemptive-repeat-identical policies. As results, the Laplace transforms associated with waiting time and response time, and the z-transform for the number of jobs in the system as well as their expectations are presented.>

Bandwidth allocation strategies with access restrictions for the integration of wideband (WB) and narrowband (NB) traffic in a broadband integrated services digital network (BISDN) are studied. Access restrictions are placed on both types of traffic to meet the grade of service required for each. The WB traffic is assumed nonqueuable, whereas the NB traffic may be queued. As an improved strategy, an access strategy which allows each type of traffic to use its bandwidth in the access-restricted region with the risk of being preempted by an arriving call for which the overflowed region is reserved is considered. Under this strategy, the preempted NB calls are queued at the head of a waiting room, and preempted WB calls are queued in another finite waiting room. The performance of these strategies is analyzed in terms of access delay for NB traffic and blocking probability for WB traffic by the matrix-geometric solution method. It is shown numerically that these strategies can easily be adapted to varying traffic loads by changing the restriction levels.>

A random token protocol that incorporates priority functions is presented. In this prioritized random token protocol, the access right to a channel is given to the user of the highest priority by giving the earliest scheduling time. The protocol is suitable for high-speed networks with frequency reconfiguration, and also for mobile networks. It can be used with and without the collision detection feature. The prioritized random token protocol is analyzed with two priority classes and the delay-throughput characteristics for each priority class are determined. It is shown that the proposed protocol yields better throughput-delay performance than the p-persistent protocol of prioritized carrier sense multiple access. In addition, numerical results obtained for various system parameters are discussed and verified by simulation.>

The authors present a vector quantization (VQ) codebook design algorithm for recognition systems based on the phoneme-level hidden Markov model (HMM). In order to improve the discrimination ability of HMMs, the VQ codewords are extracted from the HMM state segments of the recognition units. An iterative VQ codebook design method based on the LVQ2 or MKM algorithm is proposed for the unified estimation of VQ codebooks and HMM parameters. Although the codebooks designed by the proposed algorithm showed larger quantization distortions for input speech spectrum than the conventional codebooks, they always improved the recognition accuracy of the recognition systems as compared to the conventional phoneme-independent or phoneme-dependent codebooks.>

The authors propose a multichannel lightwave network with a grouping property which exhibits better channel efficiency as the traffic is more likely within groups. The network also avoids both wavelength agility and pretransmission coordination problems. Two network topologies are considered. One is a multichannel lightwave network with full connection within groups. While it requires many receivers in each network interface unit (NIU) it exhibits low queuing delay and high network capacity. The other is a multichannel lightwave network with minimal connection within groups. While it gives relatively poor delay-throughput characteristics, a network can be constructed with each NIU having only two fixed-wavelength receivers. The network capacities and the mean queuing delays for the two network topologies are presented.>

A high-speed slotted ring network with a single buffer at each station is studied. For performance analysis, the authors investigate its throughput and mean delay characteristics based on a lumped model in which all stations are aggregated at a single point on the ring with their relative positions preserved.>

The authors investigate throughput efficiencies for two basic automatic-repeat-request (ARQ) strategies, go-back-N and selective repeat disciplines, assuming that forward and backward channels have Markov error characteristics and they are mutually independent. For the two ARQ schemes, simple analytical formulas for the throughput efficiency are obtained. Particularly, in the case of the go-back-N scheme the authors show that the result of throughput analysis can be extended to a general form that includes results presented by C.H.C. Leung et al. (1988).>

Two acknowledgement (ACK) schemes for window flow control, full and partial ACK schemes, are investigated in conjunction with two error-checking schemes at every node or at the store-and-forward nodes only, in a quasi-cut-through switching network with noisy channel. It is shown that with respect to power and delay, the full ACK scheme performs better than the partial ACK scheme. On the other hand, in terms of throughput, the opposite holds. This reflects a tradeoff between throughput and delay in the contention system. Thus, the two ACK schemes can effectively be used to meet alternative user requirements for services in different application areas. The full ACK scheme yields better delay and throughput performance than that of conventional flow-controlled packet switching.>

The queuing behavior of voice and data traffic for a proposed voice/data multiplexing system is investigated. For synchronization of packetized voice in the packet-switched network, the master frame format of time-division multiplexing (TDM) is utilized. This scheme has discrete delay characteristics for packetized voice and provides a simple play-out method for voice signals. The performance of voice and data traffic is investigated by obtaining the cumulative distribution of delay of voice packets and the mean delay time of voice and data packets. The performances of the proposed system and the circuit-switched system are compared to time-assignment speech interpolation (TASI) in terms of the loss probability of voice traffic and the maximum allowable number of input voice channels for a given trunk capacity. The proposed system has been found to be more efficient than the circuit-switched system with TASI for integrating multiple voice channels in a single link. For a given trunk capacity, the system can accommodate about twice the number of input voice channels as the circuit-switched system.>

A computationally efficient codebook search method in code-excited linear prediction is proposed. The method can reduce the computational complexity by almost one half compared to the frequency-domain codebook search method that is currently regarded as the fastest search method, while giving almost the same quality of speech. This reduction is possible as a result of the simultaneous use of frequency-domain search and code vector sparsity.>

An investigation is conducted of two acknowledgment schemes for sliding window flow control in packet communications: the delayed acknowledgment (D-ACK) scheme and the window acknowledgment (W-ACK) scheme. It is shown that the acknowledgment scheme of the link layer of the X.25 protocol can be modeled by the D-ACK scheme and that of the packet layer by the W-ACK scheme. The efficiencies of these acknowledgment schemes are shown by the mean number of acknowledgment control frames generated for the transmission of a single information frame, the transmission blocking probability, and the mean transmission time of data frames. Numerical and simulation results showing that the D-ACK scheme with an appropriately selected value of predetermined waiting time is better than the W-ACK scheme are presented.>

A study is made of the performance of a class of adaptive data-driven echo cancellers (DDECs). The authors first investigate the interrelationship in a unified framework among the system equations of various data-driven echo cancellers. As a result, they obtain a new DDEC algorithm that consists of only one real-valued adaptive structure. The authors analyze and compare the convergence behavior of DDECs. They analyze their complexities when the least-mean-square (LMS) algorithm and the frequency-domain block LMS algorithm are used for adjusting the canceller weights. The results show that the echo canceller structure realized in the frequency domain has advantages in convergence rate and in implementation complexity as compared to existing DDEC structures.>

The authors propose a multiplexing frame structure that makes it possible to transmit voice messages synchronously without loss or clipping of contents. This scheme has discrete delay characteristics, and provides a simple play-out method for reproduction of voice signal. The authors investigate its performance by obtaining the cumulative distribution of delay of voice packets and the mean waiting time of data packets. It is concluded that this synchronous frame structure can easily be applied to enhance services with various transmission rates, such as flow control of message streams, node congestion control, and service-class or throughput-class negotiation of channels without significant degradation of trunk utilization.>

Voice and data transmission on a common link is modelled as a Markov renewal process. The performance of an integrated voice/data multiplexer with arbitrary message length distribution and nongated service discipline, particularly mean data message delay, is analyzed using the model. Numerical results are presented.>

The authors investigate an infinite-impulse response (IIR) adaptive digital filter (ADF) based on the nonlinear least-squares (NLS) algorithm, and compare its convergence speed to that of a self-orthogonalizing IIR ADF which is known to have fastest convergence. By simulation, it is shown that the NLS IIR ADF converges faster than other known IIR ADFs, especially for a low-order case.>

The authors propose a fast adaptive least-squares algorithm for linear-phase finite-impulse-response (FIR) filters. The algorithm requires 10m multiplications per data point where m is the filter order. This reduced computation stems from the fast adaptive forward-backward least-squares method. Both linear-phase cases i.e. with constant phase delay and with constant group delay, are examined. Simulation results demonstrate that the proposed algorithm is superior to the least-mean-squares gradient algorithm and the averaging scheme used for the modified fast Kalman algorithm.>

This letter presents an improved digital tanlock loop (DTL) based on a multi-sampling scheme. It yields extended locking range, and reduced steady-state mean and variance of phase error as compared to the conventional DTL. The performance of the first-order DTL is analyzed in the presence of white Gaussian noise, and compared to the conventional DTL.

In this letter, we propose a reduced structure for realization of a passband fractionally spaced equalizer (FSE) using the block LMS (BLMS) algorithm. Since this structure consists of one real transversal filter, its complexity is considerably simpler than existing equalizer structures. The passband FSE using the real passband input converges to the same steady-state value as that of the FSE with the complex analytic input, but its convergence time takes twice as long.

In this paper, a forward/backward adaptive digital filtering method is studied for the enhancement of noisy speech. The enhancing algorithm uses future samples as well as previous samples to estimate a current sample, and utilizes the correlational property of speech. It is effective for the enhancement of narrow-band as well as wide-band noisy speech. To improve the enhancing capability further, its modified version is also considered. This modified algorithm, of which structure is relatively simple, works for any kind of additive noise, and results in signal-to-noise ratio gain over existing enhancement algorithms by about 2 dB.

In this paper, we propose an isolated word recognition system based on the finite-state vector quantization (FSVQ) method. The recognition system can be viewed as a finite state machine composed of a codebook and next-state functions. As compared to an isolated word recognition system that uses the conventional memoryless vector quantization, the proposed system requires far less search time, and needs no segmentation of input speech, yet yields comparable recognition accuracies. For the design of next-state functions, two techniques, that is, the conditional histogram and omniscient design methods, are used, and their performances are compared in recognition of the ten Korean digits.

A typical connected digit recognition system employs dynamic programming (DP) for pattern matching. In order to reduce the computational burden of DP, we propose a frame-labeling-based data reduction technique. This approach can not only reduce input data significantly, but also provide the estimated number of digits and a sequence of candidate digits in a string for dynamic programming.

Cut-through switching is advantageous in that it can reduce the transmission delay compared with the conventional message or packet switching. In this paper, when the channel is noisy, we investigate various properties ofM/D/1quasi-cut-through switching including actual traffic intensity and overall network delay. In the analysis of delay resulting from retransmission of erroneous packets, we have included the average transmission time of negative acknowledgment signal and queueing time for the retransmitted packet so that the overall network delay can be obtained accurately. In addition, we have obtained distributions of the number of nodes to be traversed and the number of nodes through which packets pass by cut. According to the analysis results, the performance of cut-through switching is superior to that of conventional packet switching in most practical ranges of parameter values.

In this paper the performance and realization of adaptive digital filters (ADF's) for differentially coded signals have been studied. A general formula that realizes the LMS algorithm for differentially coded signals is first derived. We use this formula for realization of ADM and ADPCM ADF's. Then, applying these algorithms to channel equalization, we study the performances of the two filters for the cases of using a time-varying and a constant convergence factor. If we consider the dynamic range and the sensitivity to the variation of parameter values as the important performance factors, the ADPCM ADF with a time-varying convergence factor is preferred. However, when we consider the system complexity, the use of the ADM ADF with a time-varying convergence factor is desirable.

In this letter, we propose a frequency-weighted block LMS (FWBLMS) algorithm that is based on minimization of the frequency-weighted block mean-squared error. The FWBLMS algorithm has an important advantage over the time-domain LMS algorithm in that it can be designed with different frequency weighting on error signal depending on the relative significance of various frequency bands. Application of the FWBLMS algorithm in adaptive linear prediction of speech is discussed.

In this letter, a simple method for modulation and demodulation of minimum shift keying (MSK) is proposed. It has a serial structure, and is easier to implement than the conventional parallel type. The MSK signal is generated by a digital synthesis method and the demodulation process is done by a simple sampling scheme.

In this paper, the performance of anN-phase digital tanlock loop (DTL) in the presence of phase error disturbance is studied. TheNphase DTL is a digital loop for tracking suppressed-carrierN-ary phaseshift keyed (PSK) signals. It has linear phase characteristics in the modulo-2\pi/Nsense as a result of using the\tan^{-1}[.]function in the phase error detector. We first obtain the probability density function of phase error of the phase detector output and verify it by computer simulation. Then, we consider the optimum bandwidths of the bandpass filters in the loop that yield the minimum variance of the phase error. Also, the finite word length effect on the performance of theN-phase DTL is investigated by computer simulation. In addition, the performance of theN-phase DTL is compared to that of the digitalN-phaseI-Qloop which has sinusoidal phase characteristics. It has been found that the first-order Nphase DTL has wider lock range than the first-orderN-phaseI-Qloop in the absence of noise.

In this paper we investigate methods of improving the performance of a finite impulse response (FIR) adaptive delta modulation digital filter (ADMDF). These include modification of the ADM step size adaptation algorithm, coefficient optimization, and the reduction of computational complexity. Compared to the FIR ADMDF of Sunwoo and Un [1], the improved filter yields better performance by approximately 4 dB and requires significantly less computation time.

The performance of speech waveform coders that incorporate a noise spectral shaping filter is discussed. The coders studied are adaptive pulse code modulation (APCM), adaptive differential PCM (ADPCM), and adaptive delta modulation (ADM). For APCM and ADPCM, the noise shaping filter is designed to minimize the C-message weighted quantization noise power. As for ADM, the noise shaping filter is used to move some portion of in-band noise outside the signal band. Simulation results show that the performance improvement of these waveform coders with spectral shaping is about 0.5-3 dB over the systems without noise shaping. Although this improvement is relatively small, the waveform coders with noise spectral shaping yield subjectively more pleasing and intelligible sound than those without it.

In this paper the performances of four statistical voice/data multiplexers are analyzed and compared. The structures of these multiplexers are based on the newly proposed queueing, frame management, and integrated flow control methods. For the queueing purpose, two separate finite voice and data buffers are used to queue voice and data packets independently. For frame management, four versions of master frame management are considered. As for the integrated flow control algorithm, speech coding techniques such as variable rate coding and embedded coding are considered. To evaluate the performances of the multiplexers, analytical formulations are made from the equivalent queueing models of the proposed systems. Then, numerical results on the performance of the multiplexers are obtained by using the Gauss-Seidel iteration method. These results show that there exist tradeoffs between data and voice traffic performances for a given input traffic load and the output channel rate.

In this paper, the performances of PCM/ADPCM transcoding systems are analyzed. The coders studied are the 64 kbit/s PCM with μ-255 companding law and the 32 kbit/s ADPCM proposed by Cummiskey et al. The theoretically predicted performance agrees closely with the results of computer simulation for a wide range of input signal level. According to the results, the performance degradation resulting from the code conversion process appears to be minimal for the single tandem case. However, for the case of multiple tandem code conversions, the performance becomes significantly degraded as the number of tandem coders increases. The overall performance of the transcoding system depends largely on the performance of the coder that is inferior to the other coder being cascaded.

In this paper we analyze the convergence behavior of the frequency-domain LMS(FLMS) adaptive digital filter(ADF) using discrete Fourier transform. We obtain the optimum Wiener solution, the minimum mean-squared-error(MSE), the convergence condition, and the excess MSE in the steady state. Also, we study the convergence behavior of the FLMS ADFbased on the concept of a self-orthogonalizing algorithm. By approximating the input autocorrelation matrix as a circulant matrix, we compare the results ofthe LMS and FLMS ADF's.

Results of a performance comparison study of eight pitch extraction algorithms for noisy as well as clean speech are presented. These algorithms are the autocorrelation method with center clipping, the autocorrelation method with modified center clipping, the simplified inverse filter tracking (SIFT) method, the average magnitude difference function (AMDF) method, the pitch detection method based on LPC inverse filtering and AMDF, the data reduction method, the parallel processing method and the cepstrum method. It has been found that for pitch detection of noisy speech the algorithm that uses an AMDF or an autocorrelation function yields relatively good performance than others. A pitch detector that uses center clipped speech as an input signal is effective in pitch extraction of noisy speech. In general, preprocessing such as LPC inverse filtering or center clipping of input speech yields remarkable improvement in pitch detection.

In this paper the realization and performance of a twodimensional finite impulse response (FIR) digital filter using adaptive delta modulation (ADM) have been studied. The analytical results have been verified by computer simulation. Also, it has been shown that onedimensional DM is adequate in the realization of a two-dimensional FIR ADM digital filter. Since the ADM digital filter requires no multiplication, it offers many advantages over conventional PCM filters in cost reduction, hardware simplicity, and reduced size.

In this paper, implementation of a compact and efficient multirate speech digitizer with variable transmission rates of 2.4, 4.8, 9.6, and 14.96 kbits/s is presented. The multirate algorithm has been made based on the residual-excited linear prediction (RELP) vocoder with a transmission rate of 9.6 kbits/s. The residual encoder employed in the RELP vocoder uses hybrid companding delta modulation (HCDM). This HCDM is also used as a 14.96 kbit/s coder. If the residual in the RELP system is down-sampled before encoding, a 4.8 kbit/s coder can be realized. If the residual encoder is not used, a 2.4 kbit/s linear predictive coder (LPC) can be realized by incorporating a pitch extractor. In the 4.8 and 9.6 kbit/s coders the pitch-implanted residual excitation method has been used to generate the excitation signal to the synthesis filter. The multirate speech digitizer algorithm has been implemented using 2900 series bit-slice microprocessors. The external memory is composed of 2K RAM's and 2K ROM's. The system design is a two-bus structure with a 204 ns cycle time. With efficient hardware and software design, the multirate speech digitizer requires almost the same hardware complexity as compared with the conventional 2.4 kblt/s LPC vocoder.

In this paper we describe hardware implementation of a 2400 bit/s two-channel linear predictive vocoder. The vocoder hardware uses the latest 2900 series "bit slice" microprocessor chips, and 2K RAM's and 1K ROM's of data memory. The system design is a two bus structure with a 208 ns cycle time. A significant feature of the vocoder is that it is capable of processing two voice channels simultaneously. Throughout the paper, emphasis is placed on details of firmware development of the vocoder system. Efficient design of the vocoder hardware is also discussed.

Performance and realization of finite impulse response (FIR) digital filters that use an adaptive delta modulator as an analog/ digital converter have been studied. These filters require no multiplication and offer many advantages over conventional PCM filters including low power consumption, small size, and cost effectiveness. Analytical formulas have been derived for the expected mean-squared errors and also for the word length necessary to achieve the desired performance. Computer simulation has been done to optimize the parameter values and to verify the results of performance analysis. In addition, design of FIR ADM digital filters for processing single and multichannel signals has been considered.

In this paper four pole-zero modeling algorithms that are based on high-order pole model fitting and decomposition method have been studied, and application to robust vocoding has been considered. They are autocorrelation prediction (AP), modified Yule-Walker (MYW), modified least square (MLS) and modified least square with autocorrelation compensation (MLSAC) methods. They involve only linear equations, and therefore are computationally efficient. Among those algorithms, the MLSAC method appears to be the most effective in spectral envelope estimation of noisy as well as clean speech. According to our simulation result, the improvement resulting from the use of the MLSAC pole-zero model for noisy speech is equivalent to increasing signal-to-noise ratio (SNR) by about 5 dB when SNR of input speech is 10 dB or less. The use of the MLSAC method in multi-rate vocoding is also discussed.

Residual excited linear prediction (RELP) vocoder is considered to be a promising approach to medium rate (4.8 to 9.6 kbits/s) vocoding that does not require pitch extraction. One critical part of the RELP vocoder that affects greatly the synthetic speech quality is the spectral flattener in which high frequency components are regenerated from the baseband residual signal. In this paper several spectral flattening techniques so far proposed are first reviewed, and their advantages and disadvantages are discussed. Then, two improved methods are proposed. Those are the hybrid method and the split-band coding method. These new methods yield excellent synthetic speech quality when the transmission rate is 8 kbits/s.

In this paper, a nonuniform-sampling digital phaselocked loop (DPLL) called the digital tanlock loop (DTL), which uses a new type of phase error detector with linear phase characteristic, is studied. The main feature of the DTL is that the phase error detector, using thetan^{-1}(.)function with in-phase and quadrature samples of the incoming signal, has a linear characteristic with a period of 2π. Accordingly, the DTL can be characterized by a linear difference equation, thereby making it possible to analyze the loop easily, without approximation of nonlinearity as is usually done in analysis of a conventional DPLL with sinusoidal phase characteristic. The performances of the first- and second-order DTL's in the absence and presence of noise have been investigated by analysis and computer simulation. It is shown that the linear phase characteristic results in many attractive features in comparison to the conventional DPLL with the sinusoidal phase characteristic. These include insensitivity of the locking conditions to variation of input signal power, more noise immunity, wider lock range and less steady-state phase error of the first-order loop for an input with frequency offset, and much less sensitivity to initial phase errors in convergence of the second-order loop.

Hybrid companding delta modulation (HCDM) is known to be superior in performance to other instantaneous or syllabic companding delta modulation systems [1]. To improve its performance or to reduce the bit rate further in coding speech, we propose to use a variable-rate sampling scheme in the HCDM system. The proposed system employs several different sampling rates but transmits the output binary signal at a fixed rate using a buffer. By using the variable-rate scheme, one can improve its performance by 3 to 4 dB in signal-to-quantization noise ratio (SQNR) over the fixedrate HCDM. Detailed algorithm and computer simulation results are presented. Buffer behavior and its control are also discussed. In addition, it is shown that the performance gain of a DM system with variable-rate sampling depends on the degree of variation of the input signal.

A robust linear predictive coding (LPC) method that can be used in noisy as well as quiet environment has been studied. In this method, noise autocorrelation coefficients are first obtained and updated during non-speech periods. Then, the effect of additive noise in the input speech is removed by subtracting values of the noise autocorrelation coefficients from those of autocorrelation coefficients of corrupted speech in the course of computation of linear prediction coefficients. When signal-to-noise ratio of the input speech ranges from 0 to 10 dB, a performance improvement of about 5 dB can be gained by using this method. The proposed method is computationally very efficient and requires a small storage area.

In this paper the performance of delta modulation (DM) systems have been studied by computer simulation when the input signal to the DM coder is a voiceband data signal. First, the parameter values of three DM systems, linear DM (LDM), constant factor DM (CFDM), and continuously variable slope DM (CVSD) has been optimized for 4800 bits/s differential phase shift keying (DPSK) signal. Then, the performance of the three DM systems have been studied for ideal and noisy channels. It has been found that the peak signal-to-quantization-noise ratio (SQNR) is nearly the same regardless of coding scheme used, but CFDM yields the widest dynamic range. In a noisy channel, however, CFDM is very sensitive to channel bit errors. Considering the overall performance, CVSD appears to be the best among the three DM systems studied. Also, the performances of DM's have been compared with those of PCM and DPCM systems. In addition, we have studied the effect of DM quantization noise on modem bit error rate by the Monte Carlo simulation method. It is possible to transmit a 4800 bits/s DPSK signal at a bit error rate below 10-5by CVSD with the rate of 32 kbits/s.

We present a reasonably complete account of an improved adaptive delta modulation (ADM) system called hybrid companding delta modulation (HCDM). The HCDM system that is far superior to continuously variable slope DM (CVSD) or constant factor DM (CFDM) is advantageous, particularly for speech coding. It employs both syllabic and instantaneous companding schemes. Performance analysis of the system has been done and verified by computer simulation. In getting the mathematical formula for HCDM granular noise, a new method based on amplitude distribution is proposed. Optimization of the system parameter values by simulation is also discussed. In addition, an efficient method of hardware implementation is considered.

Three different types of pole-zero modeling have been investigated. The main concept lying in these methods is to fit a high order pole predictor to the speech spectrum, and then to decompose the resulting predictor into a pole predictor and a zero predictor. To obtain the predictor parameters by Padé approximation, either Trench algorithm or Berlekamp-Massey algorithm can be used for decomposition. In addition to Padé approximation method, the generalized inverse method and the modified Durbin's method are also considered for pole-zero modeling. Preliminary results show that the proposed methods yield better fitting of speech spectral envelope, especially of spectral nulls, than the all-pole model.

We present an improved 4800 bps LPC vocoder system that virtually eliminates the buzzy effect from synthetic speech. Excitation signal in the new system is formed by adding high-pass filtered pitch pulses or random noise to a baseband residual signal (0 - 600 Hz) that has been coded by pitch predictive PCM. Since the baseband residual is used as a part of excitation, the system is also robust to V/UV and pitch errors. According to our informal listening tests, the synthetic speech of the new system does not have the buzzy effect. As a result the vocoder speech quality is more natural than that of a conventional LPC vocoder.

New analytical formulas for quantization noise of two widely known adaptive delta modulation (ADM) systems, continuously variable slope delta modulation (CVSD) and Jayant's firstorder constant factor delta modulation (CFDM), have been obtained. These formulas are derived based upon the previously obtained results for linear delta modulation (LDM) and adaptive differential pulse code modulation (ADPCM). To verify the results, computer simulation has been done using band-limited Gaussian signals. Close agreements between the formulas and the simulation results have been obtained in the wide range of the input signal level.

A performance comparison of three representative ADM systems has been made by computer simulation using real speech. The three systems studied are continuously variable slope delta modulation (CVSD), Jayant's constant factor delta modulation (CFDM), and a modified version of Un and Magill's hybrid companding delta modulation (HCDM). Among the three systems, HCDM yields the best performance in signal-to-quantization noise ratio (SQNR) and dynamic range regardless of the channel bit error rate. Comparing CVSD and CFDM in an ideal channel, the dynamic range of the latter is significantly wider than that of the former, although their peak SQNR's are almost the same. In a noisy channel, CFDM degrades more rapidly than the other two as the bit error rate increases. In the channel with an error rate above 10-3, the use of CFDM appears to be impractical when the bit rate is below 16 khits/s. However, intelligible speech transmission is possible with HCDM or CVSD even at the error rate of as high as 10-1.

A complete algorithm of a 1200-bits/s digital formant vocoder system is described. This vocoder algorithm draws heavily on the results of recent research in linear predictive coding. The transmitting parameters are frequencies and amplitudes of the first three formants, the pitch period, voiced/unvoiced decision, and the gain. Formant bandwidths are estimated at the synthesizer by using the amplitude information. The synthesizer structure is in the parallel form. The synthetic speech quality at 1200 bits/s is reasonably good; most of the speech is intelligible and speaker-recognizable.

In this paper we present a new vocoder called the residual-excited linear prediction (RELP) vocoder. The concept of the RELP vocoder combines the advantages of linear predictive coding (LPC) and voice-excited vocoding. In the RELP system, vocal tract modeling is done by the LPC technique, and the LPC residual signal is used as the excitation signal. After low-pass filtering the residual signal is coded by adaptive delta modulation and is spectrally flattened before being fed in the LPC synthesizer. The range of the transmission rate is typically between 6 and 9.6 kbits/s; the synthetic speech in this range is quite good. As the transmission rate is lowered, the synthetic speech quality degrades very gradually. Since no pitch extraction is required, the vocoder is robust in any operating environment.

Based on the Fokker-Planck techniques, the time-dependent mean and variance of phase error of the first-order phase-locked loop in the presence of white Gaussian noise have been obtained via stochastic differential equations. The steady-state variance reduced from the transient case is simple in its form and agrees closely with the previous results obtained by different approaches.