The Dempster-Shafer (DS) belief theory constitutes a powerful framework for modeling and reasoning with a wide variety of uncertainties due to its greater expressiveness and flexibility. As in the Bayesian probability theory, the DS theoretic (DST) conditional plays a pivotal role in DST strategies for evidence updating and fusion. However, a major limitation in employing the DST framework in practical implementations is the absence of an efficient and feasible computational framework to overcome the prohibitive computational burden DST operations entail. The work in this article addresses the pressing need for efficient DST conditional computation via the novel computational model DS-Conditional-All. It requires significantly less time and space complexity for computing the Dempster's conditional and the Fagin-Halpern conditional, the two most widely utilized DST conditional strategies. It also provides deeper insight into the DST conditional itself, and thus acts as a valuable tool for visualizing and analyzing the conditional computation. We provide a thorough analysis and experimental validation of the utility, efficiency, and implementation of the proposed data structure and algorithms. A new computational library, which we refer to as DS-Conditional-One and DS-Conditional-All (DS-COCA), is developed and harnessed in the simulations.

Adequate representative statistical training data needed for machine learning algorithms are often unavailable and, when available, they are often mired in incomplete/missing data. Imputation of such data must be guided by the relationships among different variables and/or by data `missingness' mechanisms. Interval-valued (IV) probabilities are better suited in situations where such information is unavailable. We take the viewpoint that IV probabilities (IVPs) emerge from a single underlying probability distribution about which one has only partial information. PrBounds, the IVPs that this vantage point engenders, offer a fresh perspective of the IV counterpart notions of conditioning and independence and enable reasoning to be carried out in much the same manner as one would with probabilities. When the attribute values are unknown/missing or are known to lie within a set of values, PrBounds can be efficiently learnt by a frequency counting method. The probabilities associated with an arbitrary imputation strategy, including the underlying `true' probabilities, are guaranteed to lie within the PrBounds learnt in this manner. We present an experiment to illustrate the proposed framework.

The problem of inferring the hidden state of individual nodes in social/sensor networks in which node activities affect their neighbors is growing in importance. We present an undirected generative model, a type of probabilistic model that has so far not been used for modeling latent variables influenced by neighbors in a network. We also propose an efficient inference method based on variational inference principles which, in contrast to sampling methods used in most existing models, is scalable to larger networks. While training is intractable in general, by using stochastic methods to approximate the intractable derivative, we show that our model can be trained using the maximum likelihood method by formulating the model as an exponential family distribution. The results demonstrate that the proposed undirected model can accurately infer latent states compared to baseline methods.

We examine the design of self-organizing mobile adaptive networks with multiple targets in which the network nodes form distinct clusters to learn about and purse multiple targets, all while moving in a cohesive collision-free manner. We build upon previous distributed diffusion-based adaptive learning networks that focused on a single target to examine the case with multiple targets in which the nodes do not know the number of targets, and exchange local information with their neighbors in their learning objectives. In particular, we design a method allowing the nodes to switch the target they are tracking thereby engendering the formation of distinct stable learning groups that can split up and purse their distinct targets over time. We provide analytical mean stability and steady state mean-square deviation results along with simulations that demonstrate the efficacy of the proposed method.

In linear prediction of speech, the 1-norm error minimization criterion has been shown to provide a valid alternative to the 2-norm minimization criterion. However, unlike 2-norm minimization, 1-norm minimization does not guarantee the stability of the corresponding all-pole filter and can generate saturations when this is used to synthesize speech. In this paper, we introduce two new methods to obtain intrinsically stable predictors with the 1-norm minimization. The first method is based on constraining the roots of the predictor to lie within the unit circle by reducing the numerical range of the shift operator associated with the particular prediction problem considered. The second method uses the alternative Cauchy bound to impose a convex constraint on the predictor in the 1-norm error minimization. These methods are compared with two existing methods: the Burg method, based on the 1-norm minimization of the forward and backward prediction error, and the iteratively reweighted 2-norm minimization known to converge to the 1-norm minimization with an appropriate selection of weights. The evaluation gives proof of the effectiveness of the new methods, performing as well as unconstrained 1-norm based linear prediction for modeling and coding of speech.

In a multi-agent data fusion scenario, agents may iteratively exchange their states to arrive at a consensus state which signifies ‘general agreement’ among the agents. Agent states that are being exchanged may have been generated from hard (i.e., physics based) or soft (i.e., human based evidence. such as opinions or beliefs regarding an event) sensors. Convergence analysis becomes an extremely challenging problem in such complex fusion environments, which may involve communication delays, ad-hoc paths, etc. In this paper, we analyze consensus of a Dempster-Shafer theoretic (DST) fusion operator by formulating the consensus problem as finding common fixed points of a pool of paracontracting operators. Due to its DST basis, this consensus protocol can deal with a wider variety of data imperfections characteristic of hard+soft data fusion environments. It also easily adapts itself to networks where agent states are captured with probability mass functions because they can be considered a special case of DST models.

Dempster-Shafer (DS) belief theory provides a convenient framework for the development of powerful data fusion engines by allowing for a convenient representation of a wide variety of data imperfections. The recent work on the DS theoretic (DST) conditional approach, which is based on the Fagin-Halpern (FH) DST conditionals, appears to demonstrate the suitability of DS theory for incorporating both soft (generated by human-based sensors) and hard (generated by physics-based sources) evidence into the fusion process. However, the computation of the FH conditionals imposes a significant computational burden. One reason for this is the difficulty in identifying the FH conditional core, i.e., the set of propositions receiving nonzero support after conditioning. The conditional core theorem (CCT) in this paper redresses this shortcoming by explicitly identifying the conditional focal elements with no recourse to numerical computations, thereby providing a complete characterization of the conditional core. In addition, we derive explicit results to identify those conditioning propositions that may have generated a given conditional core. This "converse" to the CCT is of significant practical value for studying the sensitivity of the updated knowledge base with respect to the evidence received. Based on the CCT, we also develop an algorithm to efficiently compute the conditional masses (generated by FH conditionals), provide bounds on its computational complexity, and employ extensive simulations to analyze its behavior.

The aim of this paper is to provide an overview of Sparse Linear Prediction, a set of speech processing tools created by introducing sparsity constraints into the linear prediction framework. These tools have shown to be effective in several issues related to modeling and coding of speech signals. For speech analysis, we provide predictors that are accurate in modeling the speech production process and overcome problems related to traditional linear prediction. In particular, the predictors obtained offer a more effective decoupling of the vocal tract transfer function and its underlying excitation, making it a very efficient method for the analysis of voiced speech. For speech coding, we provide predictors that shape the residual according to the characteristics of the sparse encoding techniques resulting in more straightforward coding strategies. Furthermore, encouraged by the promising application of compressed sensing in signal compression, we investigate its formulation and application to sparse linear predictive coding. The proposed estimators are all solutions to convex optimization problems, which can be solved efficiently and reliably using, e.g., interior-point methods. Extensive experimental results are provided to support the effectiveness of the proposed methods, showing the improvements over traditional linear prediction in both speech analysis and coding.

In many contexts, one is confronted with the problem of extract ing information from large amounts of different types soft data (e.g., text) and hard data (from e.g., physics-based sensing systems). In handling hard data, signal and data processing offers a wealth of methods related to modeling, estimation, tracking, and inference tasks. However, soft data present several challenges that necessitate the development of new data processing methods. For example, with suitable statistical natural language processing (NLP) methods, text can be converted into logic statements that are associated with various forms of associated uncertainty related to the credibility of the statement, the reliability of the text source, and so forth. In combining or fusing soft data with either soft or hard data, one must deploy methods that can suitably preserve and update the uncertainty associated with the data, thereby providing uncertainty bounds related to any inferences regarding semantics. Since standard Bayesian probabilistic approaches have problems with suitably handling uncertain logic statements, there is an emerging need for new methods for processing heterogeneous data. In this paper, we describe a framework for fusing soft and hard data based on the Dempster-Shafer (DS) belief theoretic approach which is well-suited to the task of capturing the types of models and uncertain rules that are more typical of soft data. Since the effectiveness of traditional DS methods has been hampered by high computational requirements, we base the processing framework on our new conditional approach to DS theoretic evidence updating and fusion. We address the issue of laying the foundation for a theoretically justifiable, and computationally efficient framework for fusing soft and hard data taking into account the inherent data uncertainty such as reliability and credibility. Moreover, we present an illustrative ex ample that highlights the potential for the DS conditional approach for fusing heterogeneous data.

We consider linear predictive coding and noise shaping for coding and transmission of auto-regressive (AR) sources over lossy networks. We generalize an existing framework to arbitrary filter orders and propose use of fixed-lag smoothing at the decoder, in order to further reduce the impact of transmission failures. We show that fixed-lag smoothing up to a certain delay can be obtained without additional computational complexity by exploiting the state-space structure. We prove that the proposed smoothing strategy strictly improves performance under quite general conditions. Finally, we provide simulations on AR sources, and channels with correlated losses, and show that substantial improvements are possible.

Linear prediction of speech based on 1-norm minimization has already proved to be an interesting alternative to 2-norm minimization. In particular, choosing the 1-norm as a convex relaxation of the 0-norm, the corresponding linear prediction model offers a sparser residual better suited for coding applications. In this paper, we propose a new speech modeling technique based on reweighted 1-norm minimization. The purpose of the reweighted scheme is to overcome the mismatch between 0-norm minimization and 1-norm minimization while keeping the problem solvable with convex estimation tools. Experimental results prove the effectiveness of the reweighted 1-norm minimization, offering better coding properties compared to 1-norm minimization.

In this paper, we describe a new approach to cope with packet loss in speech coders. The idea is to split the information present in each speech packet into two components, one to independently decode the given speech frame and one to enhance it by exploiting inter-frame dependencies. The scheme is based on sparse linear prediction and a redefinition of the analysis-by-synthesis process. We present Mean Opinion Scores for the presented coder with different degrees of packet loss and show that it performs similarly to frame dependent coders for low packet loss probability and similarly to frame independent coders for high packet loss probability. We also present ideas on how to make the coder work synergistically with the channel loss estimate.

Distributed Source Coding (DSC) has been widely studied in applications such as video coding and distributed sensor networks. However, DSC has not been widely explored for low delay and low bit rate applications such as quantization of speech Line Spectral Frequencies (LSFs). This is due to the difficulty of modeling and analyzing the effects of imperfect side information resulting from the previous packet losses, quantization noise and decoding errors. In this paper, we present methods for modeling, analyzing and designing Wyner-Ziv(WZ) quantizers for jointly Gaussian vector data with imperfect side information. In particular, we show the decomposition of the quantizer design problem for the vector data into independent scalar design subproblems. Then we demonstrate the analytical techniques to compute the optimum step size and bit allocation for each scalar dimension to minimize the decoder expected Mean Squared Error(MSE). The simulation results verify the analytical results obtained in this paper.

Encouraged by the promising application of compressed sensing in signal compression, we investigate its formulation and application in the context of speech coding based on sparse linear prediction. In particular, a compressed sensing method can be devised to compute a sparse approximation of speech in the residual domain when sparse linear prediction is involved. We compare the method of computing a sparse prediction residual with the optimal technique based on an exhaustive search of the possible nonzero locations and the well known Multi-Pulse Excitation, the first encoding technique to introduce the sparsity concept in speech coding. Experimental results demonstrate the potential of compressed sensing in speech coding techniques, offering high perceptual quality with a very sparse approximated prediction residual.

Active Queue Management (AQM) methods attempt to predict and control network router queue levels and provide feedback regarding network congestion to data sources through packet marking/ dropping. AQM methods have not employed statistical signal processing principles largely due to the requirement of low complexity. In this paper, we apply optimal filtering and target tracking methods to the design of AQM. In particular, we develop Kalman Filter based AQM which results in router queues with reduced queue level variance. To account for networks with more bursty traffic, we use Interacting Multiple Models (IMM) which similarly result in reduced queue variance in simulations with both long-term and bursty short-term traffic. In comparisons with other AQM methods, these low complexity target tracking-based AQM methods give a more constant queue length without any loss in source throughput.

Gaussian mixture model (GMM)-based Kalman predictive coders have been shown to perform better than baseline GMM recursive coders in predictive coding of line spectral frequencies (LSFs) for both clean and packet loss conditions However, these stationary GMM Kalman predictive coders were not specifically designed for operation in packet loss conditions. In this paper, we demonstrate an approach to the the design of GMM-based predictive coding for packet loss channels. In particular, we show how a stationary GMM Kalman predictive coder can be modified to obtain a set of encoding and decoding modes, each with different Kalman gains. This approach leads to more robust performance of predictive coding of LSFs in packet loss conditions, as the coder mismatch between the encoder and decoder are minimized. Simulation results show that this Robust GMM Kalman predictive coder performs better than other baseline GMM predictive coders with no increase in complexity. To the best of our knowledge, no previous work has specifically examined the design of GMM predictive coders for packet loss conditions.

Gaussian mixture model (GMM)-based predictive coding of line spectral frequencies (LSFs) has gained wide acceptance. In such coders, each mixture of a GMM can be interpreted as defining a linear predictive transform coder. In this paper, we use Kalman filtering principles to model each of these linear predictive transform coders to present GMM Kalman predictive coding. In particular, we show how suitable modeling of quantization noise leads to an adaptive a posteriori GMM that defines a signal-adaptive predictive coder that provides improved coding of LSFs in comparison with the baseline recursive GMM predictive coder. Moreover, we show how running the GMM Kalman predictive coders to convergence can be used to design a stationary GMM Kalman predictive coding system which again provides improved coding of LSFs but now with only a modest increase in run-time complexity over the baseline. In packet loss conditions, this stationary GMM Kalman predictive coder provides much better performance than the recursive GMM predictive coder, and in fact has comparable mean performance to a memoryless GMM coder. Finally, we illustrate how one can utilize Kalman filtering principles to design a postfilter which enhances decoded vectors from a recursive GMM predictive coder without any modifications to the encoding process.

In a multisensor target tracking application running on a shared network, at what bit rates should the sensors send their measurements to the tracking fusion center? Clearly, the sensors cannot use arbitrary rates in a shared network, and a standard network rate control algorithm may not provide rates amenable to effective target tracking. For Kalman filter-based multisensor target tracking, we derive a utility function that captures the tracking quality of service as a function of the sensor bit rates. We incorporate this utility function into a network rate resource allocation framework, deriving a distributed rate control algorithm for a shared network that is suitable for current best effort packet networks, such as the Internet. In simulation studies, the new rate control algorithm engenders significantly better tracking performance than a standard rate control method, while the ordinary data transfer flows continue to effectively operate while using their standard rate control methods.

Recently developed cooperative protocol with distributed path selection provides a simple and practical means of achieving full cooperative diversity in wireless networks. While the best path selection method can significantly improve bit error rate (BER) performance, it may cause unequal power consumption among relay nodes, which may reduce the lifetime of energy-constrained networks. A path selection method under the equal power constraint has been developed for the amplifyand- forward (AF) protocol, but there is no such method for the decode-and-forward (DF) protocol. In this paper, we develop a distributed path selection method with an equal power constraint for the DF protocol. We also analyze the BER performance of our path-selection method. Numerical results demonstrate that the proposed method can guarantee equal power consumption, while achieving full diversity as the best path selection method and providing significant performance gain relative to noncooperative communication.

In a multi-sensor target tracking application running on a shared network, at what bit-rates should the sensors send their measurements to the tracking fusion center? Clearly, the sensors cannot use arbitrary rates in a shared network, and a standard network rate control algorithm may not provide rates amenable to effective target tracking. For Kalman Filter-based multi-sensor target tracking, we derive a utility function that captures the tracking Quality of Service as a function of bit-rate. We incorporate this utility function into a network rate resource allocation framework, deriving a distributed rate control algorithm for a shared network that does not require network re-design. In simulation studies, the new rate-control algorithm engenders much better tracking performance than a standard rate-control method.

Gaussian Mixture Model (GMM)-based predictive coding of line spectral frequencies (lsf’s) has gained wide acceptance. In such coders, each mixture of a GMM can be interpreted as defining a linear predictive transform coder. In this paper we optimize each of these linear predictive transform coders using Kalman predictive coding techniques to present GMM Kalman predictive coding. In particular, we show how suitable modeling of quantization noise leads to an adaptive a-posteriori GMM that defines a signal-adaptive predictive coder that provides superior coding of lsfs in comparison with the baseline GMM predictive coder. Moreover, we show how running the Kalman predictive coders to convergence can be used to design a stationary predictive coding system which again provides superior coding of lsfs but now with no increase in run-time complexity over the baseline.

All-pole spectral envelope estimates based on linear prediction (LP) for speech signals often exhibit unnaturally sharp peaks, especially for high-pitch speakers. In this paper, regularization is used to penalize rapid changes in the spectral envelope, which improves the spectral envelope estimate. Based on extensive experimental evidence, we conclude that regularized linear prediction outperforms bandwidth-expanded linear prediction. The regularization approach gives lower spectral distortion on average, and fewer outliers, while maintaining a very low computational complexity.

Depending upon the type of feedback that is utilized, network congestion control schemes can be classified into two categories: marking/loss based (e.g., TCP Reno) or delay based (e.g., TCP Vegas). Delay-based schemes have garnered much attention due to their higher network throughput than loss-based methods. Delay provides a much finer-grained measure of congestion than packet loss or packet marking feedback. However, when there are multiple bottleneck links or inadequate buffer sizes in the path between a source and destination, delay information alone is insufficient for revealing the incipient network congestion. In this paper, we consider the design of a congestion control scheme that transcends the two normal categories and instead jointly exploits both delay and marking/loss feedback. In particular, we introduce the concept of normalized queuing delay, and demonstrate how a controller based on joint feedback provides high throughput even under highly dynamic network conditions as evidenced by ns2 simulations.

A well-known problem with linear prediction is that its estimate of the spectral envelope often has sharp peaks for high-pitch speakers. These peaks are anomalies resulting from contamination of the spectral envelope by the spectral fine structure. We investigate the method of regularized linear prediction to find a better estimate of the spectral envelope and compare the method to the commonly used approach of bandwidth expansion. We present simulations over voiced frames of female speakers from the TIMIT database, where the envelope modeling accuracy is measured using a log spectral distortion measure. We also investigate the coding properties of the methods. The results indicate that the new regularized LP method is superior to bandwidth expansion, with an insignificant increase in computational complexity

The Internet low bit-rate coder (iLBC) is now widely used for voice over Internet protocol (VoIP) applications. Unlike speech coders such as those based on code excited linear prediction (CELP), the iLBC achieves superior robustness to packet loss by avoiding inter-frame coding dependencies. While robustness to packet loss is essential, a VoIP codec should also possess the flexibility to change its source coding rate in order to counter network congestion and facilitate joint source channel coding for wireless channels. Previously, we presented a new variation of the iLBC encoding procedure which yielded a more efficient, rate-flexible result. In an effort to improve performance at lower source rates, we present various improvements to the original framework. Specifically, we reallocate bits from the adaptive codebook procedure; reduce the length of the start state vector; utilize an adaptive pulse gain quantization scheme; and extend the use of entropy coding. Overall, the various combined improvements result in the modified iLBC (with entropy coding) achieving a rate reduction of 2.0 to 2.9 kbps when compared to the original fixed-rate iLBC without any loss in quality. In comparisons with adaptive multi-rate (AMR), the modified iLBC coder remarkably exhibits equivalent perceptual evaluation of speech quality (PESQ) scores as the AMR coder at 10.2 and 12.2 kbps, and out-performs AMR for all packet loss rates. This is a significant result as the modified iLBC performs equivalent to AMR without exploiting inter-frame redundancies

Packet loss concealment (PLC) at a receiver has a substantial effect on the speech quality in voice over IP. Most conventional PLC systems have largely relied upon variations of signal repetition and overlap-add interpolation which can produce speech signals that do not follow the larger overall statistical trends. in this paper, we demonstrate how hidden Markov models can be utilized to effect PLC based on statistical signal processing. In particular, we show how HMM-based PLC yields conditional density functions that can be utilized by various statistical estimation methods that produce signal parameter estimates that produce more natural variation than conventional PLC methods, thereby providing much better speech quality

We address the problem of optimizing resource allocation for Perceptually based unequal packet protection (PUPP) in a packet based voice carrying network. For that purpose, we design a novel real-time working perceptually based classifier (PBC) optimizing the assignment of voice packets to either a premium (Pch) or an ordinary (Och) transmission channel with regard to packet perceptual importance. In particular, our PBC is based on sliding window optimization (SWO) and implement PESQa, an improved method to real-time estimation of speech quality. Based on this PBC and a differentiated service (DS) implementation of the Pch/Och, objective results indicate that 70% premium packet assignments optimizes performance over a broad range of loss scenarios on a bottleneck link. Additionally, packet loss statistics gives a clear indication on criteria for optimizing PUPP Pch/Och

As voice over IP proliferates, packet loss concealment (PLC) at the receiver has emerged as an important factor in determining voice quality of service. Through the use of heuristic variations of signal and parameter repetition and overlap-add interpolation to handle packet loss, conventional PLC systems largely ignore the dynamics of the statistical evolution of the speech signal, possibly leading to perceptually annoying artifacts. To address this problem, we propose the use of hidden Markov models for PLC. With a hidden Markov model (HMM) tracking the evolution of speech signal parameters, we demonstrate how PLC is performed within a statistical signal processing framework. Moreover, we show how the HMM is used to index a specially designed PLC module for the particular signal context, leading to signal-contingent PLC. Simulation examples, objective tests, and subjective listening tests are provided showing the ability of an HMM-based PLC built with a sinusoidal analysis/synthesis model to provide better loss concealment than a conventional PLC based on the same sinusoidal model for all types of speech signals, including onsets and signal transitions

The Internet low bit-rate coder (iLBC) has emerged as a candidate for VoIP applications. iLBC is able to achieve superior robustness to packet loss. A VoIP codec should also possess the agility to adjust its source coding rate in order to react to network congestion and to be amenable to joint source channel coding for wireless channels. Towards this end, we develop a new formulation of the iLBC encoding process that allows for a variable rate iLBC. In particular, we demonstrate how the LP excitation signal is constructed from a much shorter vector of 'start state' samples through a non-square synthesis matrix that captures the effects of the adaptive codebook operations. With this new framework, the search and quantization of the start state is re-formulated as an analysis by synthesis matching problem. We demonstrate how a multi-pulse (MP) approach can be utilized to effect a variable rate coding solution for this new framework. A variable rate coder with the MP start state achieves better performance than the adaptive multi-rate (AMR) coder at 12.2 and 10.2 kbps for packet loss rates greater than 4 %.

TCP congestion avoidance mechanisms determine methods by which a source adjusts its window size according to network conditions. Although network calculus has been utilized to study window flow control, the use of network calculus to determine an optimal window controller and to provide analytical guidance to TCP congestion avoidance has persisted as an open problem. For the first time within a network calculus setting, we determine an optimal window size control method for general flow control problems. We also show that the basic TCP congestion avoidance mechanisms in TCP Vegas, enhanced TCP Vegas and fast TCP can be viewed as different approaches to approximating the optimal NC controller, with each TCP variant making different assumptions in terms of parameter estimation and control implementation strategy. Therefore, the network calculus controller reveals the inherent underlying structure in TCP congestion avoidance. Furthermore, we demonstrate through ns-2 simulations that an approximation of a particular NC controller achieves performance gains in terms of link throughput and source node throughput fairness with respect to TCP Vegas, enhanced TCP Vegas and fast TCP.

In voice over IP, typical forward error correction (FEC) schemes to combat packet loss allocate an equal amount of error-control resources to each voice packet, regardless of the perceptual importance of a packet. Recognizing the unequal perceptual importance of voice packets, we propose signal-adaptive unequal error protection methods in which certain packets are allocated more error-control resources than others. In particular, the amount of error protection provided to a packet is determined through an analysis by the expected decoder synthesis paradigm ensconced within a rate-distortion Lagrangian optimization framework. Therefore, the sender evaluates various protection policies by anticipating the behavior of the decoder's packet loss concealment (PLC) algorithm for various loss event probabilities. In this manner, perceptually critical voice packets that cannot be easily replaced by a PLC are provided with greater error protection. For a given average bit-rate, a simple unequal error protection scheme provides a 0.2 to 0.3 advantage in PESQ-MOS (perceptual evaluation of speech quality mean opinion score) over the conventional equal error control schemes.

In many VoIP systems, the end-to-end voice QoS is largely dictated by the packet loss rate and the packet loss concealment (PLC) method at the receiver. Typical PLC algorithms are usually based on simple signal extrapolation methods that can produce perceptually annoying artifacts. We present a taxonomy of PLC artifacts that degrade voice quality, and demonstrate how this knowledge can be used to define a PLC-driven labeling of critical voice packets. In particular, we show how the packets whose loss will produce an annoying artifact can be labeled for transmission over a virtual premium channel effected through either a DiffServe method, or adaptive FEC on a best-effort ordinary network. With a PLC-driven labeling of critical packets transmitted over a premium channel, the VoIP application can achieve large potential QoS gains over random labeling approaches, and is more robust to packet loss over an ordinary channel. To automate the labeling of packets, we demonstrate the feasibility of a sender-based packet classifier to detect the packets whose loss will produce PLC artifacts.

We present all-pole models based upon the minimum variance distortionless response (MVDR) spectrum for spectral modeling of speech. The MVDR method, which is popular in array processing, provides all-pole spectra that are robust for modeling both voiced and unvoiced speech. Although linear prediction (LP) is a popular method for obtaining all-pole model parameters, LP spectral envelopes overestimate and overemphasize the medium and high pitch voiced speech spectral powers, thereby featuring unwanted sharp contours, and do not improve in spectral envelope modeling performance as the filter order is increased. In contrast, the MVDR all-pole spectrum which can be easily obtained from the LP coefficients, features improved spectral envelope modeling as the filter order is increased. In particular, the high order MVDR spectrum models voiced speech spectra very well, particularly at the perceptually important harmonics, and features a smooth contoured envelope. Furthermore, the MVDR spectrum can be based upon either conventional time domain correlation estimates or upon spectral samples, a task that is common in frequency domain speech coding. In particular, the MVDR spectrum of sufficient order provides an all-pole envelope that models a set of spectral samples exactly. In addition, the MVDR all-pole spectrum is also suitable for modeling unvoiced speech spectra.

We present several analytical properties of minimum variance distortionless response (MVDR) based all-pole models that demonstrate the advantages and usefulness of these models for speech spectral coding. In particular, we show that a sufficient order MVDR all-pole model provides a spectral envelope that fits a set of spectral samples exactly with a parameterization convenient for quantization purposes. In addition, we show that MVDR all-pole filters provide a monotonically decreasing spectral distortion with increasing filter order. Furthermore, we show that the MVDR all-pole filter possesses the flexibility to be obtained from correlations based upon either spectral samples or conventional time-domain correlations. Finally, exploiting the insight gained from MVDR modeling, we introduce a novel class of constrained all-pole models for efficient spectral coding. In this approach, a subset of the line spectral frequency (LSF) parameters associated with the all-pole model are judiciously fixed, leading to a simpler model parameterization.

In this paper, we propose some new modeling techniques that provide a more synergistic approach to multistage time-domain speech compression. In particular, we propose a new error criterion for determining all-pole filters, and a unique method for jointly coding the pulse information in excitation vectors. The new error criterion for determining all-pole filters is based upon minimizing the sum of the residual signal's absolute values raised to a power less than one. It is shown to be a desirable cost function for yielding residual signals that are more sparse, and consequently better suited for multistage compression than linear prediction residuals. Statistical reasons supporting the new criterion are also provided. Furthermore, exploiting the properties of, and the relationship between, the linear prediction and minimum variance spectra, we propose a novel parameter set for jointly coding the excitation vector's pulse position, sign, and gain information.

In this paper we propose the MVDR method, which is based upon the minimum variance distortionless response (MVDR) spectrum estimation method, for modeling voiced speech. Developed to overcome some of the shortcomings of linear prediction models, the MVDR method provides better models for medium and high pitch voiced speech. The MVDR model is an all-pole model whose spectrum is easily obtained from a modest non-iterative computation involving the linear prediction coefficients thereby retaining some of the computational attractiveness of LPC methods. With the proper choice of filter order, which is dependent on the number of harmonics, the MVDR spectrum models the formants and spectral powers of voiced speech exactly. An efficient reduced model order MVDR method is developed to further enhance its applicability. An extension of the reduced order MVDR method for recovering the correct amplitudes of the harmonics of voiced speech is also presented.