Blind multiplicative watermarking schemes for speech signals using wavelets and discrete cosine transform are presented. Watermarked signals are modeled using a generalized Gaussian distribution (GGD) and Cauchy probability model. Detectors are developed employing generalized likelihood ratio test (GLRT) and locally most powerful (LMP) approach. The LMP scheme is used for the Cauchy distribution, while the GLRT estimates the gain factor as an unknown parameter in the GGD model. The detectors are tested using Monte Carlo simulation and results show the superiority of the proposed LMP/Cauchy detector in some experiments

Parameter-embedded watermarking is effected through slight perturbations of parametric models of some deeply-integrated dynamics of the speech signal. In set-membership filtering (SMF) based parametric watermarking, linear predictor (LP) coefficients of the original speech are modified subject to an objective fidelity constraint. SMF is used to obtain sets of allowable parameter perturbations (i.e., watermarks) subject to a constraint on the error between the watermarked and original material. This paper discusses the robustness of SMF based watermarking to filtering, quantization and combination attacks. An important consideration in watermark robustness is the energy of the watermark signal (difference between watermarked and original signals). The most robust watermark is obtained from perturbed LP coefficients at the boundary of the membership set. A constrained optimization problem is solved to obtain the best watermarks for filtering and quantization attacks

This paper describes a new class of multi-pulse coders based on bounded-error identification. Multi-pulse coders provide effective representations of speech signals by extracting linear prediction parameters and approximating the effect of excitation signals with strategic sparse impulse sequences. These sparse impulses can be derived according to a set-membership optimization set volume reduction strategy. The proposed method offers the added benefit of improved scalar quantization through the choice of coder parameters from a feasibility set

Known forms of Wiener filter do not adequately address the discrete-time signal processing of continuous random processes. In this paper, we first formulate this problem and subsequently derive its solution. The crux of the work is the development of an autocorrelation function invariant discrete equivalent of a continuous system. This in turn enables us to transform the hybrid Wiener filtering problem into a purely discrete one. An example demonstrates the concepts presented, and draws a comparison of ACF-invariance with other discretization schemes.

Parametric watermarking is effected by modifying the linear predictor coefficients of speech. In this work, the parameter noise is analyzed when watermarked speech is subjected to additive white and colored noise in the time domain. The paper presents two detection techniques for parametric watermarking. The first approach uses the Neyman-Pearson criterion to solve a binary decision problem. In the second approach, discriminant functions based on the minimum-error-rate criterion are used to determine which one of the many watermarks was embedded or if no watermark is present. Experiments with speech data are used to determine the false-alarm and missed detection rates

An analytical method for design and performance analysis of language models (LM) is described, and an example interactive software tool based on the technique is demonstrated. The LM performance analysis does not require on-line simulation or experimentation with the recognition system in which the LM is to employed. By exploiting parallels with signal detection theory, a profile of the LM as a function of the design parameters is given in a set of curves analogous to a receiver-operating-characteristic display.

Advances in formulating spoken document retrieval for a new National Gallery of the Spoken Word (NGSW) are addressed. NGSW is the first large-scale repository of its kind, consisting of speeches, news broadcasts, and recordings from the 20th century. After presenting an overview of the audio stream content of the NGSW, with sample audio files from U.S. Presidents from 1893 to the present, an overall system diagram is proposed with a discussion of critical tasks associated with effective audio information retrieval. These include advanced audio segmentation, speech recognition model adaptation for acoustic background noise and speaker variability, and information retrieval using natural language processing for text query requests that include document and query expansion. For segmentation, a new evaluation criterion entitled fused error score (FES) is proposed, followed by application of the CompSeg segmentation scheme on DARPA Hub4 Broadcast News (30.5% relative improvement in FES) and NGSW data. Transcript generation is demonstrated for a six-decade portion of the NGSW corpus. Novel model adaptation using structure maximum likelihood eigenspace mapping shows a relative 21.7% improvement. Issues regarding copyright assessment and metadata construction are also addressed for the purposes of a sustainable audio collection of this magnitude. Advanced parameter-embedded watermarking is proposed with evaluations showing robustness to correlated noise attacks. Our experimental online system entitled "SpeechFind" is presented, which allows for audio retrieval from a portion of the NGSW corpus. Finally, a number of research challenges such as language modeling and lexicon for changing time periods, speaker trait and identification tracking, as well as new directions, are discussed in order to address the overall task of robust phrase searching in unrestricted audio corpora.

Embedded speech processing systems require stringent memory allocation a,nd computing resources. To minimize such resources, a simple, flexible HMM evaluation technique is presented which employs a state-space formulation in conjunction with a simplified likelihood measure. The method offers several advantages including the ability to reduce redundant computation and memory allocation across models, and a flexible structure that can exploit known results concerning statespace systems. Although performance is insignificantly effected in preliminary experiments, these benefits are achieved at the cost of a weaker coupling between the two stochastic processes that define the HMM. We augment the method with a Markov chain model of the ohservations to compensate for the weaker state coupling. Preliminary experiments are used Lo analyze recognition performance and as a basis for discussion. 1.

Phone tokenization followed by n-gram language modeling has consistently provided good results for the task of language identification. In this paper, this technique is generalized by using Gaussian mixture models as the basis for tokenizing. Performance results are presented for a system employing a GMM tokenizer in conjunction with multiple language processing and score combination techniques. On the 1996 CallFriend LID evaluation set, a 12-way closed set error rate of 17% was obtained.

A general formulation for speech watermarking through parametric modeling is suggested, then the paper focuses on a watermarking technique based on linear-predictive (LP) modeling of speech. In the particular strategy employed here, information is embedded by modifying the autocorrelation values of the original speech. The amount of information that can be embedded is subject to fidelity constraints. The modified LP coe#cients derived from the new set of autocorrelation values are used for reconstructing the watermarked speech. The perceptual quality of the watermarked speech depends on the relative energy of the embedded watermark, the watermark sequence used, and the LP model order. Robustness of the technique to various signal processing operations and attacks like compression, cropping, and additive noise are studied via experiments on a small speech application. Factors a#ecting watermark robustness and related security issues are also discussed.

Published results indicate that automatic language identification (LID) systems that rely on multiple-language phone recognition and n-gram language modeling produce the best performance in formal LID evaluations. By contrast, Gaussian mixture model (GMM) systems, which measure acoustic characteristics, are far more efficient computationally but have tended to provide inferior levels of performance. This paper describes two GMM-based approaches to language identification that use shifted delta cepstra (SDC) feature vectors to achieve LID performance comparable to that of the best phone-based systems. The approaches include both acoustic scoring and a recently developed GMM tokenization system that is based on a variation of phonetic recognition and language modeling. System performance is evaluated on both the CallFriend and OGI corpora. 1.

Most digital watermarking techniques are susceptible to damage by data cropping. Although the effects of cropping might not be perceptible, watermark recovery may be rendered difficult or impossible due to the desynchronization of the recovery process. The transform encryption coding (TEC) based watermarking algorithm was presented at ICASSP 2000 (see Ruiz, F.J. and Deller, J.R., Jr, IEEE International Conference on Acoustics, Speech and Signal Processing 2000). This paper investigates the performance of TEC watermarking in the presence of cropping, and presents an algorithm that identifies cropped samples and recovers watermarks from the damaged record. Implementation details and experimental results under different environmental conditions are presented.

Optimal bounding ellipsoid (OBE) identification algorithms are noted for their simplicity and ability to leverage model error-bound knowledge for improved parameter convergence. However, the OBE convergence rate is dependent on the pointwise "tightness" of the model error-bound estimates. Since the least upper bound on the model error is often unknown, the convergence rate is compromised by the need to overestimate error-bounds lest the integrity of the process be violated by underestimation. We present an effective under-bounding safeguard against system model violations in OBE processing. Simulation examples in state estimation and speech processing demonstrate the efficacy of the under-bounding safeguard.

A novel speech watermarking technique has been developed in support of the National Gallery of the Spoken Word (NGSW), a Digital Libraries Initiative, Phase II, sponsored project. Transform encryption coding (TEC) is used to compress and encrypt both the watermark and the coversignal. TEC is employed in combination with a masking algorithm which takes into account the signal energy of the watermark and the coversignal. The process of creating the stegosignal further encrypts the watermark. This dual encryption results in a high level of protection for original materials as the two proper keys (quasi m-arrays) are required to remove each watermark. An application to a sample utterance from the NGSW is presented.

Optimal bounding ellipsoid (OBE) algorithms offer an attractive alternative to traditional least squares methods for identifying linear-in-parameters signal and system models due to their low computational efficiency, superior tracking ability, and selective updating that permits processor sharing among tasks. These benefits are further enhanced by multiweight optimization (MWO) which yields improved per-point parameter convergence. This paper introduces the MWO process and describes advances in its implementation including the incorporation of a forgetting factor for improved tracking, a new method for efficient weight computation, and extensions to volume-minimizing OBE algorithms. Simulation studies illustrate the results.

Application of set-membership (SM) identification to real-time speech processing is made possible by the optimal bounding ellipsoid algorithm with automatic bound estimation (OPE-APE) that blindly deduces model-input bounds. To date, lack of any tenable approach to estimating bounds in speech models has rendered these interesting new SM methods impractical. OBE-ABE is consistently convergent, offers significant computational advantages, and provides a set of feasible solutions in finite time.

Set-membership algorithms, including the conventional optimal bounding ellipsoid (OBE) algorithm, require a priori knowledge of exact error bounds which is unknown in most applications. Conservative (over-estimated) error bounds used in practice lead to inconsistent parameter estimation. The novel OBE algorithm with automatic bound estimation (OBE-ABE) is shown to be consistently convergent without a priori knowledge of error bounds, even in correlated-error environments. Computationally efficient variants of this algorithm for both time-invariant and time-varying systems are presented. Simulations are performed to demonstrate the merit of the algorithms.

Practical application of optimal bounding ellipsoid identification is made possible by the optimal bounding ellipsoid algorithm with automatic bound estimation (OBE-ABE), which automatically estimates model error bounds. Lack of tenable bounds in many real problems has rendered these interesting new methods impractical. OBE-ABE is consistently convergent under proven conditions, and offers significant computational advantages.

Partitioned search (PS) of a language model results in dramatic complexity reduction, and improved recognition rates and noise robustness. PS is applied to recognition of continuous-speech taken from the TIMIT database. Comparisons of the results using the following approaches for recognizing continuous speech are made; a quiet environment, and when the speech is corrupted by intermittent noise: 1. PS, 2. left-to-right search using random selection of as many vertices as in the PS cases for evaluation, and 3. conventional left-to-right stack search with pruning.>

A new training method for feedforward neural networks is presented which exploits results from matrix perturbation theory for significant training time improvement. This theory is used to assess the effect of a particular training pattern on the weight estimates prior to its inclusion in any iteration. Data which do not significantly change the weights are not used in that iteration obviating the computation expense of updating.>

Set-membership (SM) identification, which refers to a class of algorithms using certain a priori knowledge about a parametric model to constrain the solutions to certain sets, is considered. The focus is on a class of SM-based techniques that are of particular interest in applications requiring real-time processing. The optimal bounding ellipsoid (OBE) algorithms are interpreted as a blending of the classical least-square error minimization approach with knowledge of bounds on model errors arising from SM considerations. Using this interpretation, a general framework embracing all currently used OBE algorithms is developed, and strategies for adaptation and for implementation on parallel machines are discussed. Computational complexity benefits are considered for the various algorithms. The treatment is tutorial, leaving many of the formal details to an appendix that presents an archival theoretical treatment of the key results. A second appendix gives an overview of current research in the general SM identification field.>

Redundant computations occur when a set of hidden Markov models (HMMs) is evaluated with respect to an observation string. A formal restructuring of the HMM allows the redundancy to be identified and removed. An O((1- kappa )NT) complexity HMM results, with the compression index kappa in

The convergence and bias properties of a general class of optimal bounding ellipsoid (OBE) algorithms are discussed. OBE algorithms are set-membership (SM) based identification algorithms which are applied to models which are linear-in-parameters, and are closely related to weighted recursive least square error (WRLS) methods.>

A strategy is developed which can be applied to any version, adaptive or non-adaptive, of the set membership weighted recursive least squares (SM-WRLS) algorithm to improve the computational efficiency. A significant reduction in computational complexity can be achieved by employing a suboptimal test for information content in the incoming data. The main issue is to avoid the computations of an O(m/sup 2/) checking procedure, where m is the number of parameters to be estimated, which is required to check for the existence of useful data. Since most of the time these computations result in the rejection of incoming data, a more efficient test which reduces the complexity of the algorithm to O(m) is presented.>

A self-organizing neural network is presented which automatically learns the number and type of spectral features from speech examples. The learning algorithm is analyzed with respect to its convergence and stability properties. The “strength” of the presence of the learned features is registered by the network to effect recognition of further speech presentations. The network consists of two layers of feature detectors, each layer of which is self-organized, and the outputs of the second layer are time-aligned in the present design using dynamic time warping. The significance of the two-layer structure, as well as general architectural advantages of the network, are discussed. Results of experiments involving various isolated word recognition tasks, including single and multi-speaker training and recognition, and the recognition of speech of a nonverbal individual, are reported.

An adaptive set membership identification algorithm with a very flexible forgetting scheme is presented. In preliminary experiments, the method yields highly accurate estimates using very few of the data, and quickly adapts to fast-changing dynamics. A compact systolic architecture to implement this algorithm is developed which uses O(m) cells and reduces the computational complexity to O(m) operations per observation, where m represents the number of parameters to be estimated in a linear system or signal model.>

It is shown that the optimal bounding ellipsoid (OBE) algorithm for identifying an ARMAX system can be formulated as a conventional weighted recursive least squares estimator with special weights. In this framework the OBE can be implemented using contemporary algorithms developed for least squares solutions on systolic machines. An example of a systolic processor for OBE is given, and computational complexity issues are considered.>

A conventional HMM (hidden Markov model) method has been applied to the problem of isolated-word cerebral palsy speech recognition. A full-structure HMM was found to provide best results because of the conditions of high variability and small amounts of training data. To overcome the inadequacies of the conventional method, an enhanced clipping procedure has been developed which aids in the removal of variability in both the training and recognition phases of the HMM procedure. The performance of isolated-word recognition was significantly improved when this enhanced procedure was applied in a case study.>

An intelligent communication device is being developed to assist the nonverbal, motor disabled in the generation of written and spoken messages. The device is based on a knowledge base of the grammatical rules and message elements. A 'belief' reasoning scheme makes use of highly uncertain speech-recognition information in combination with the embedded knowledge to optimize the process of message search. The authors focus on the integration of that speech information. Experimental testing of a prototype system to assess the effects of speech evidence is discussed.>

A frame control strategy is presented which directs an efficient graph search when information available for finding a path is distributed throughout the graph. The frame structure requires (worst case) only O(K/sup N/) memory size instead of O(K/sup 2N/), which is needed for a matrix representation of an N-level, K-branch tree graph. An efficiency analysis is detailed, and its application to a communication system for the nonverbal, severely motor disabled is discussed.>

A method for recognition of speech of the nonverbal is presented. Results of the application of the procedure to simple utterances by cerebral palsied individuals and implications for an overall communications device are addressed.

The theory of set membership (SM) identification is formulated, and applied to linear prediction (LP) analysis of speech. The LP parameters of a simulated vowel are identified as an illustration. The SM strategy results in a significant computational savings due to rejection of data which are informationless in the SM sense.

A new digital signal processing technique has been developed for the evaluation of laryngeal function based on features extracted from an accurate estimate of the glottal volume velocity waveform. Analysis of a preliminary group of pathologic and normal subjects indicates that the resulting features can be used to effectively classify voice samples into normal and abnormal types. In addition, analysis of cases in the space of features indicates potential for differential diagnosis of disease or, at least, indication of severity of the dysfunction.

An autoregressive (AR) systemx{n} = \Sigma\min{i=1}\max{M} a_{i}(n) x(n-i) + A(n) w(n)where{a_{i}(n)}, and A(n) are general stationary stochastic processes and w(n) is white noise, is well modeled by a similar equation in which the random processes are replaced by their averages, if the resulting constant coefficient model is sufficiently lowpass, low gain (LPLG) [1]. By computer simulation, we investigate the performance of this approximation for various AR systems. It is found that "sufficiently LPLG" implies a very restricted class (very LPLG) for the general case of unknown (large) perturbations on{a_{i}(n), A(n)}. Only as the perturbations become small can the very LPLG restriction be relaxed.