Dysarthria is a motor speech disorder often characterized by reduced speech intelligibility through slow, uncoordinated control of speech production muscles. Automatic Speech recognition (ASR) systems may help dysarthric talkers communicate more effectively. To have robust dysarthria-specific ASR, sufficient training speech is required, which is not readily available. Recent advances in Text-To-Speech (TTS) synthesis multi-speaker end-to-end systems suggest the possibility of using synthesis for data augmentation. In this paper, we aim to improve multi-speaker end-to-end TTS systems to synthesize dysarthric speech for improved training of a dysarthria-specific DNN-HMM ASR. In the synthesized speech, we add dysarthria severity level and pause insertion mechanisms to other control parameters such as pitch, energy, and duration. Results show that a DNN-HMM model trained on additional synthetic dysarthric speech achieves WER improvement of 12.2% compared to the baseline, the addition of the severity level and pause insertion controls decrease WER by 6.5%, showing the effectiveness of adding these parameters. Audio samples are available at https://mohammadelc.github.io/SpeechGroupUKY/

This paper compares the difference in articulation patterns between native (L1) and non-native (L2) Mandarin speakers of English, for the purpose of providing an understanding of mispronunciation behaviors of L2 learners. Consensus transcriptions from the Electromagnetic Articulography Mandarin Accented English (EMA-MAE) corpus are used to identify commonly occurring substitution errors for consonants and vowels. Phoneme level alignments of the utterances produced by speech recognition models are used to extract articulatory feature vectors representing correct and substituted sounds from L1 and L2 speaker groups respectively. The articulatory features that are significantly different between the two groups are identified along with the direction of error for the L2 speaker group. Experimental results provide information about which types of substitutions are most common and which specific articulators are the most significant contributors to those errors.

Long short-term memory (LSTM) networks have been widely used in automatic speech recognition (ASR). This paper proposes a novel dynamic temporal residual learning mechanism for LSTM networks to better explore temporal dependencies in sequential data. The temporal residual learning mechanism is implemented by applying shortcut connections with dynamic weights to temporally adjacent LSTM outputs. Two types of dynamic weight generation methods are proposed: using a secondary network and using a random weight generator. Experimental results on Wall Street Journal (WSJ) speech recognition dataset reveal that our proposed methods have surpassed the baseline LSTM network.

We propose an innovative approach to enhance welding operations by using a cyber-physical system (CPS) with layered architecture and enabling a robot to be effectively operated by its commanding human. This article focuses on the recognition of the commanding human's intention that should be executed by the robot. To this end, a virtual reality (VR) system based on the HTC Vive is used to create a remote virtual welding environment. Human hand movement speed data is collected and used to train a hidden Markov model (HMM) using the Baum-Welch algorithm. The Bayesian information criterion (BIC) is applied to determine the number of hidden states. The state occupancy probability distribution (the probability of each state at a given time) is estimated based on the human hand movement speed sequence using the forward algorithm. The human intention, defined as the intended movement in this article, is then estimated as the statistical expectation of the observable variables. Using the proposed human intention estimation algorithm, the intended movement recognized from the raw movement data is smoother, which is preferable in welding tasks. A 6-DoF industrial robot, UR-5 with a custom gas tungsten arc welding (GTAW) torch installed, works as the final performer of the welding jobs. The robot receives the intended movement data from the HMM and uses this to assist the human welding operators. Welding experiments have been conducted both with and without the proposed human intention recognition (IR) algorithm. The results show that the robot can help the operators complete welding tasks with better performance using the proposed IR system, supporting the effectiveness of the proposed VR robot-assisted welding system. Note to Practitioners-Welding is not only labor-intensive but also requires real-time adaption to the process, which is challenging for robots/machines but relatively straightforward for humans. Using a human commanded robot to perform welding can liberate humans from laborious operations and hazardous environments. To this end, a virtual reality (VR) system is used to create a virtual welding environment for a human to view the process remotely and for a human to pass his/her resultant adaptation to the robot through hand movements. However, hand movements do not always fully represent the intended adaptation of the commanding human, and the recognition of such human intention is fundamental in such a proposed method. This article established the mathematical framework for the recognition of the human intention and, thus, the foundation for effective assistance of robots to humans.

Speaker embeddings achieve promising results on many speaker verification tasks.Phonetic information, as an important component of speech, is rarely considered in the extraction of speaker embeddings.In this paper, we introduce phonetic information to the speaker embedding extraction based on the x-vector architecture.Two methods using phonetic vectors and multi-task learning are proposed.On the Fisher dataset, our best system outperforms the original x-vector approach by 20% in EER, and by 15%, 15% in minDCF08 and minDCF10, respectively.Experiments conducted on NIST SRE10 further demonstrate the effectiveness of the proposed methods.

Linear discriminant analysis-probabilistic linear discriminant analysis (LDA-PLDA) is a standard and effective backend in the field of speaker verification. The object of LDA is to perform dimensionality reduction while minimizing within-class covariance and maximizing between-class covariance. For a target class (or speaker), our task is to make a binary decision about whether a test utterance is from a specific target speaker. Generally, the nontarget test utterances that are close to the target speaker are easily misjudged. Inspired by this idea, we propose a local pairwise linear discriminant analysis (LPLDA) algorithm. This new method focuses on maximizing the local pairwise covariance, which represents the local structure between the target class samples and neighboring nontarget class samples, instead of the between-class covariance, which represents the global structure of the data. Experiments on the NIST SRE 2010, 2014, and 2016 database show that, the proposed LPLDA-PLDA backend has significant performance improvements over the LDA-PLDA backend.

Text-dependent speaker verification is becoming popular in the speaker recognition society. However, the conventional i-vector framework which has been successful for speaker identification and other similar tasks works relatively poorly in this task. Researchers have proposed several new methods to improve performance, but it is still unclear that which model is the best choice, especially when the pass-phrases are prompted during enrollment and test. In this paper, we introduce four modeling methods and compare their performance on the newly published RedDots dataset. To further explore the influence of different frame alignments, Viterbi and forward-backward algorithms are both used in the HMM-based models. Several bottleneck features are also investigated. Our experiments show that, by explicitly modeling the lexical content, the HMM-based modeling achieves good results in the fixed-phrase condition. In the prompted-phrase condition, GMM-HMM and i-vector/HMM are not as successful. In both conditions, the forward-backward algorithm brings more benefits to the i-vector/HMM system. Additionally, we also find that even though bottleneck features perform well for text-independent speaker verification, they do not outperform MFCCs on the most challenging Imposter-Correct trials on RedDots.

Acoustic-to-articulatory inversion, the estimation of articulatory kinematics from an acoustic waveform, is a challenging but important problem. Accurate estimation of articulatory movements has the potential for significant impact on our understanding of speech production, on our capacity to assess and treat pathologies in a clinical setting, and on speech technologies such as computer aided pronunciation assessment and audio-video synthesis. However, because of the complex and speaker-specific relationship between articulation and acoustics, existing approaches for inversion do not generalize well across speakers. As acquiring speaker-specific kinematic data for training is not feasible in many practical applications, this remains an important and open problem. This paper proposes a novel approach to acoustic-to-articulatory inversion, Parallel Reference Speaker Weighting (PRSW), which requires no kinematic data for the target speaker and a small amount of acoustic adaptation data. PRSW hypothesizes that acoustic and kinematic similarities are correlated and uses speaker-adapted articulatory models derived from acoustically derived weights. The system was assessed using a 20-speaker data set of synchronous acoustic and Electromagnetic Articulography (EMA) kinematic data. Results demonstrate that by restricting the reference group to a subset consisting of speakers with strong individual speaker-dependent inversion performance, the PRSW method is able to attain kinematic-independent acoustic-to-articulatory inversion performance nearly matching that of the speaker-dependent model, with an average correlation of 0.62 versus 0.63. This indicates that given a sufficiently complete and appropriately selected reference speaker set for adaptation, it is possible to create effective articulatory models without kinematic training data.

Articulatory data such as that collected via electromagnetic articulography (EMA) are valuable for many speech applications, including speech modeling, recognition, and synthesis. Nearly all current EMA applications and methods focus on the use of positional sensor data, even though modern 3D-EMA systems also capture sensor orientation, which provides significant additional information about articulator posture and vocal tract shape. To address this problem, this paper introduces a new method for interpolating untracked tongue fleshpoint positions from the combined position and orientation data of three EMA sensors on the tongue, using additional reference sensors for evaluating interpolation error. Comparison of interpolated and measured data illustrated effectiveness of the new method in providing additional tongue shape and position features. The results suggest that analytic methods that combine sensor position and orientation data are able to improve the characterization of tongue kinematics even using a small number of EMA sensors.

Protection from spoofing attacks is an essential component of speaker verification systems. This paper proposes a novel approach to detect such attacks by utilizing supervectors derived from spectral magnitude and phase information. Three countermeasures are chosen to represent these important information. To combine different countermeasures, score fusion and an antispoofing supervector (ASSV) are used. Experiments conducted on ASVspoof 2015 show that the combination of magnitude and phase information obtains relative 90% improvement in terms of the equal error rate (EER) compared to the best subsystem in the development set. The two systems can also be fused to further improve the performance. In addition to accuracy improvements, the new supervector framework is extensible and allows for a more flexible interface to the back-end classifier design.

Sensorimotor adaptation is an important focus in the study of motor learning for non-disordered speech, but has yet to be studied substantially for speech rehabilitation. Speech adaptation is typically elicited experimentally using LPC resynthesis to modify the sounds that a speaker hears himself producing. This method requires that the participant be able to produce a robust speech-acoustic signal and is therefore not well-suited for talkers with dysarthria. We have developed a novel technique using electromagnetic articulography (EMA) to drive an articulatory synthesizer. The acoustic output of the articulatory synthesizer can be perturbed experimentally to study auditory feedback effects on sensorimotor learning. This work aims to compare sensorimotor adaptation effects using our articulatory resynthesis method with effects from an established, acoustic-only method. Results suggest that the articulatory resynthesis method can elicit speech adaptation, but that the articulatory effects of the two methods differ.

There is a significant need for more comprehensive electromagnetic articulography (EMA) datasets that can provide matched acoustics and articulatory kinematic data with good spatial and temporal resolution. The Marquette University Electromagnetic Articulography Mandarin Accented English (EMA-MAE) corpus provides kinematic and acoustic data from 40 gender and dialect balanced speakers representing 20 Midwestern standard American English L1 speakers and 20 Mandarin Accented English (MAE) L2 speakers, half Beijing region dialect and half are Shanghai region dialect. Three dimensional EMA data were collected at a 400 Hz sampling rate using the NDI Wave system, with articulatory sensors on the midsagittal lips, lower incisors, tongue blade and dorsum, plus lateral lip corner and tongue body. Sensors provide three-dimensional position data as well as two-dimensional orientation data representing the orientation of the sensor plane. Data have been corrected for head movement relative to a fixed reference sensor and also adjusted using a biteplate calibration system to place the data in an articulatory working space relative to each subject's individual midsagittal and maxillary occlusal planes. Speech materials include isolated words chosen to focus on specific contrasts between the English and Mandarin languages, as well as sentences and paragraphs for continuous speech, totaling approximately 45 minutes of data per subject. A beta version of the EMA-MAE corpus is now available, and the full corpus is in preparation for public release to help advance research in areas such as pronunciation modeling, acoustic-articulatory inversion, L1-L2 comparisons, pronunciation error detection, and accent modification training.

This paper introduces the use of three physiologically-motivated features for speaker identification, Residual Phase Cepstrum Coefficients (RPCC), Glottal Flow Cepstrum Coefficients (GLFCC) and Teager Phase Cepstrum Coefficients (TPCC). These features capture speaker-discriminative characteristics from different aspects of glottal source excitation patterns. The proposed physiologically-driven features give better results with lower model complexities, and also provide complementary information that can improve overall system performance even for larger amounts of data. Results on speaker identification using the YOHO corpus demonstrate that these physiologically-driven features are both more accurate than and complementary to traditional mel-frequency cepstral coefficients (MFCC). In particular, the incorporation of the proposed glottal source features offers significant overall improvement to the robustness and accuracy of speaker identification tasks.

Speech sensorimotor adaptation is the short-term learning of modified articulator movements evoked through sensory-feedback perturbations. A common experimental method manipulates acoustic parameters, such as formant frequencies, using real time resynthesis of the participant's speech to perturb auditory feedback. While some studies have examined phrases comprised of vowels, diphthongs, and semivowels, the bulk of research on auditory feedback-driven sensorimotor adaptation has focused on vowels in neutral contexts (/hVd/). The current study investigates coarticulatory influences of adjacent consonants on sensorimotor adaptation. The purpose is to evaluate differences in the adaptation effects for vowels in consonant environments that vary by place and manner of articulation. In particular, we addressed the hypothesis that contexts with greater intra-articulator coarticulation and more static articulatory postures (alveolars and fricatives) offer greater resistance to vowel adaptation than contexts with primarily inter-articulator coarticulation and more dynamic articulatory patterns (bilabials and stops). Participants completed formant perturbation-driven vowel adaptation experiments for varying CVCs. Results from discrete formant measures at the vowel midpoint were generally consistent with the hypothesis. Analyses of more complete formant trajectories suggest that adaptation can also (or alternatively) influence formant onsets, offsets, and transitions, resulting in complex formant pattern changes that may reflect modifications to consonant articulation

Broadening our understanding of the components and processes of speech sensorimotor learning is crucial to furthering methods of speech neurorehabilitation. Recent research in limb sensorimotor control has used virtual environments to study learning in novel sensorimotor working spaces. Comparable experimental paradigms have yet to be undertaken to study speech learning. We present acoustic and kinematic data obtained from participants producing vowels in unfamiliar articulatory-acoustic working spaces using a virtual vocal tract. Talkers with dysarthria and healthy controls were asked to produce vowels using an electromagnetic articulograph-driven speech synthesizer for participant-controlled auditory feedback. The aim of the work was to characterize performance within and between groups to generate hypotheses regarding experimental manipulations that may bolster our understanding of speech sensorimotor learning. Results indicate that dysarthric talkers displayed relatively reduced acoustic working spaces and somewhat more variable acoustic targets compared to controls. Kinematic measures of articulatory dynamics, particularly peak speed and movement jerk-cost, were idiosyncratic and did not dissociate talker groups. These findings suggest that individuals with dysarthria and healthy talkers may use idiosyncratic movement strategies in learning to control a virtual vocal tract, but that dysarthric talkers may nonetheless exhibit acoustic limitations that parallel deficits in speech intelligibility.

The selection of effective articulatory features is an important component of tasks such as acoustic-to-articulator inversion and articulatory synthesis. Although it is common to use direct articulatory sensor measurements as feature variables, this approach fails to incorporate important physiological information such as palate height and shape and thus is not as representative of vocal tract cross section as desired. We introduce a set of articulator feature variables that are palate referenced and normalized with respect to the articulatory working space in order to improve the quality of the vocal tract representation. These features include normalized horizontal positions plus the normalized palatal height of two midsagittal and one lateral tongue sensor, as well as normalized lip separation and lip protrusion. The quality of the feature representation is evaluated subjectively by comparing the variances and vowel separation in the working space and quantitatively through measurement of acoustic-to-articulator inversion error. Results indicate that the palate-referenced features have reduced variance and increased separation between vowels spaces and substantially lower inversion error than direct sensor measures.

In this study, the authors propose multichannel weighted Euclidean (WE) and weighted cosh (WCOSH) cost function estimators for speech enhancement in the distributed microphone scenario. The goal of the work is to illustrate the advantages of utilising additional microphones and modified cost functions for improving signal‐to‐noise ratio (SNR) and segmental SNR (SSNR) along with log‐likelihood ratio (LLR) and perceptual evaluation of speech quality (PESQ) objective metrics over the corresponding single‐channel baseline estimators. As with their single‐channel counterparts, the perceptually‐motivated multichannel WE and WCOSH estimators are functions of a weighting law parameter, which influences attention of the noisy spectral amplitude through a spectral gain function, emphasises spectral peak (formant) information, and accounts for auditory masking effects. Based on the simulation results, the multichannel WE and WCOSH cost function estimators produced gains in SSNR improvement, LLR output and PESQ output over the single‐channel baseline results and unweighted cost functions with the best improvements occurring with negative values of the weighting law parameter across all input SNR levels and noise types.

Speaker identification and verification has received a great deal of attention from the speech community, and significant gains in robustness and accuracy have been obtained over the past decade [1], [2]. However, the features used for identification are still primarily representations of overall spectral characteristics, and thus the models are primarily phonetic in nature, differentiating speakers based on overall pronunciation patterns. This creates difficulties in terms of the amount of enrollment data and complexity of the models required to cover the phonetic space, especially in tasks such as cross-lingual verification where enrollment and testing data may not have similar phonetic coverage. This paper introduces the use of a new feature for speaker verification, residual phase cepstral coefficients (RPCC), to capture speaker characteristics from their vocal excitation patterns. Results on a cross-lingual speaker verification task taken from the NIST 2004 SRE demonstrate that these RPCC features are significantly more accurate than traditional mel-frequency cepstral coefficients (MFCC) when the amount of enrollment data available for training is limited. Additionally, because of the significant differences in the nature of the features, combining MFCC and RPCC features shows an improvement in verification results over MFCCs alone. Index Terms: speaker verification, glottal source excitation, residual phase cepstrum, GMM, UBM.

In this paper, the authors propose the Beta-Order Minimum Mean-Square Error (MMSE) Spectral Amplitude estimator with Chi statistical models for the speech priors. The new estimator incorporates both a shape parameter on the distribution and cost function parameter. The performance of the MMSE Beta-Order Spectral Amplitude estimator with Chi speech prior is evaluated using the Segmental Signal-to-Noise Ratio (SSNR) and Perceptual Evaluation of Speech Quality (PESQ) objective quality measures. From the experimental results, the new estimator provides gains of 0-3 dB and 0-0.3 in SSNR and PESQ improvements over the corresponding MMSE Beta-Order MMSE Spectral Amplitude estimator with the standard Rayleigh statistical models for the speech prior. Index Terms: speech enhancement, amplitude estimation, phase estimation, parameter estimation

In this paper, a new hierarchical Bayesian speaker adaptation method called HMAP is proposed that combines the advantages of three conventional algorithms, maximum a posteriori (MAP), maximum-likelihood linear regression (MLLR), and eigenvoice, resulting in excellent performance across a wide range of adaptation conditions. The new method efficiently utilizes intra-speaker and inter-speaker correlation information through modeling phone and speaker subspaces in a consistent hierarchical Bayesian way. The phone variations for a specific speaker are assumed to be located in a low-dimensional subspace. The phone coordinate, which is shared among different speakers, implicitly contains the intra-speaker correlation information. For a specific speaker, the phone variation, represented by speaker-dependent eigenphones, are concatenated into a supervector. The eigenphone supervector space is also a low dimensional speaker subspace, which contains inter-speaker correlation information. Using principal component analysis (PCA), a new hierarchical probabilistic model for the generation of the speech observations is obtained. Speaker adaptation based on the new hierarchical model is derived using the maximum a posteriori criterion in a top-down manner. Both batch adaptation and online adaptation schemes are proposed. With tuned parameters, the new method can handle varying amounts of adaptation data automatically and efficiently. Experimental results on a Mandarin Chinese continuous speech recognition task show good performance under all testing conditions.

The shifted delta cepstrum (SDC) is a widely used feature extraction for language recognition (LRE). With a high context width due to incorporation of multiple frames, SDC outperforms traditional delta and acceleration feature vectors. However, it also introduces correlation into the concatenated feature vector, which increases redundancy and may degrade the performance of backend classifiers. In this paper, we first propose a time-frequency cepstral (TFC) feature vector, which is obtained by performing a temporal discrete cosine transform (DCT) on the cepstrum matrix and selecting the transformed elements in a zigzag scan order. Beyond this, we increase discriminability through a heteroscedastic linear discriminant analysis (HLDA) on the full cepstrum matrix. By utilizing block diagonal matrix constraints, the large HLDA problem is then reduced to several smaller HLDA problems, creating a block diagonal HLDA (BDHLDA) algorithm which has much lower computational complexity. The BDHLDA method is finally extended to the GMM domain, using the simpler TFC features during re-estimation to provide significantly improved computation speed. Experiments on NIST 2003 and 2007 LRE evaluation corpora show that TFC is more effective than SDC, and that the GMM-based BDHLDA results in lower equal error rate (EER) and minimum average cost (Cavg) than either TFC or SDC approaches.

This paper demonstrates a speech enhancement system based on an efficient auditory coding approach, coding of time-relative structure using spikes. The spike coding method can more compactly represent the non-stationary characteristics of speech signals than the Fourier transform or wavelet transform. Enhancement is accomplished through the use of MMSE thresholding on the spike code. Experimental results show that compared with the spectral domain logSTSA filter, both the subjective spectrogram evaluation and objective SSNR improvement for the proposed approach is better in suppressing noise in high noise situations, with fewer musical artifacts.

This paper presents a minimum mean-square error spectral phase estimator for speech enhancement in the distributed multiple microphone scenario. The estimator uses Gaussian models for both the speech and noise priors under the assumption of a diffuse incoherent noise field representing ambient noise in a widely dispersed microphone configuration. Experiments demonstrate significant benefits of using the optimal multichannel phase estimator as compared to the noisy phase of a reference channel.

This paper describes unsupervised speech/speaker cluster validity measures based on a dissimilarity metric, for the purpose of estimating the number of clusters in a speech data set as well as assessing the consistency of the clustering procedure. The number of clusters is estimated by minimizing the cross-data dissimilarity values, while algorithm consistency is evaluated by calculating the dissimilarity values across multiple experimental runs. The method is demonstrated on the task of Beluga whale vocalization clustering.

In this paper, we propose an MMSE a priori SNR estimator for speech enhancement. This estimator has similar benefits to the well-known decision-directed approach, but does not require an ad-hoc weighting factor to balance the past a priori SNR and current ML SNR estimate with smoothing across frames. Performance is evaluated in terms of estimation error and segmental SNR using the standard logSTSA speech enhancement method. Experimental results show that, in contrast with the decision-directed estimator and ML estimator, the proposed SNR estimator can help enhancement algorithms preserve more weak speech information and efficiently suppress musical noise.

In this paper, a new method for statistical estimation of Mel-frequency cepstral coefficients (MFCCs) in noisy speech signals is proposed. Previous research has shown that model-based feature domain enhancement of speech signals for use in robust speech recognition can improve recognition accuracy significantly. These methods, which typically work in the log spectral or cepstral domain, must face the high complexity of distortion models caused by the nonlinear interaction of speech and noise in these domains. In this paper, an additive cepstral distortion model (ACDM) is developed, and used with a minimum mean-squared error (MMSE) estimator for recovery of MFCC features corrupted by additive noise. The proposed ACDM-MMSE estimation algorithm is evaluated on the Aurora2 database, and is shown to provide significant improvement in word recognition accuracy over the baseline.

In this paper, we evaluate the use of appended jitter and shimmer speech features for the classification of human speaking styles and of animal vocalization arousal levels. Jitter and shimmer features are extracted from the fundamental frequency contour and added to baseline spectral features, specifically Mel-frequency cepstral coefficients (MFCCs) for human speech and Greenwood function cepstral coefficients (GFCCs) for animal vocalizations. Hidden Markov models (HMMs) with Gaussian mixture models (GMMs) state distributions are used for classification. The appended jitter and shimmer features result in an increase in classification accuracy for several illustrative datasets, including the SUSAS dataset for human speaking styles as well as vocalizations labeled by arousal level for African elephant and Rhesus monkey species.

This paper introduces the Greenwood Function Cepstral Coefficient (GFCC) and Generalized Perceptual Linear Prediction (GPLP) feature extraction models for the analysis of animal vocalizations across arbitrary species. These features are generalizations of the well-known Mel-Frequency Cepstral Coefficient (MFCC) and Perceptual Linear Prediction (PLP) approaches, tailored to take optimal advantage of available knowledge of each species*auditory frequency range and/or audiogram data. Illustrative results are presented comparing use of the GFCC and GPLP features versus MFCC features over the same frequency ranges.

The ability of a standard hidden Markov model (HMM) or expanded state HMM (ESHMM) to accurately model duration distributions of phonemes is compared with specific duration-focused approaches such as semi-Markov models or variable transition probabilities. It is demonstrated that either a three-state ESHMM or a standard HMM with an increased number of states is capable of closely matching both Gamma distributions and duration distributions of phonemes from the TIMIT corpus, as measured by Bhattacharyya distance to the true distributions. Standard HMMs are easily implemented with off-the-shelf tools, whereas duration models require substantial algorithmic development and have higher computational costs when implemented, suggesting that a simple adjustment to HMM topologies is perhaps a more efficient solution to the problem of duration than more complex approaches.

This paper introduces a novel time-domain approach to modeling and classifying speech phoneme waveforms. The approach is based on statistical models of reconstructed phase spaces, which offer significant theoretical benefits as representations that are known to be topologically equivalent to the state dynamics of the underlying production system. The lag and dimension parameters of the reconstruction process for speech are examined in detail, comparing common estimation heuristics for these parameters with corresponding maximum likelihood recognition accuracy over the TIMIT data set. Overall accuracies are compared with a Mel-frequency cepstral baseline system across five different phonetic classes within TIMIT, and a composite classifier using both cepstral and phase space features is developed. Results indicate that although the accuracy of the phase space approach by itself is still currently below that of baseline cepstral methods, a combined approach is capable of increasing speaker independent phoneme accuracy.

A novel method for speech recognition is presented, utilizing nonlinear/chaotic signal processing techniques to extract time-domain based, reconstructed phase space features. This work examines the incorporation of trajectory information into this model as well as the combination of both MFCC and RPS feature sets into one joint feature vector. The results demonstrate that integration of trajectory information increases the recognition accuracy of the typical RPS feature set, and when MFCC and RPS feature sets are combined, improvement is made over the baseline. This result suggests that the features extracted using these nonlinear techniques contain different discriminatory information than the features extracted from linear approaches alone.

A new signal classification approach is presented that is based upon modeling the dynamics of a system as they are captured in a reconstructed phase space. The modeling is done using full covariance Gaussian mixture models of time domain signatures, in contrast with current and previous work in signal classification that is typically focused on either linear systems analysis using frequency content or simple nonlinear machine learning models such as artificial neural networks. The proposed approach has strong theoretical foundations based on dynamical systems and topological theorems, resulting in a signal reconstruction, which is asymptotically guaranteed to be a complete representation of the underlying system, given properly chosen parameters. The algorithm automatically calculates these parameters to form appropriate reconstructed phase spaces, requiring only the number of mixtures, the signals, and their class labels as input. Three separate data sets are used for validation, including motor current simulations, electrocardiogram recordings, and speech waveforms. The results show that the proposed method is robust across these diverse domains, significantly outperforming the time delay neural network used as a baseline.

This paper presents a novel application of speech processing research, classification of African elephant vocalizations. Speaker identification and call classification experiments are performed on data collected from captive African elephants in a naturalistic environment. The features used for classification are 12 mel-frequency cepstral coefficients plus log energy computed using a shifted filter bank to emphasize the infrasound range of the frequency spectrum used by African elephants. Initial classification accuracies of 83.8% for call classification and 88.1% for speaker identification were obtained. The long-term goal of this research is to develop a universal analysis framework and robust feature set for animal vocalizations that can be applied to many species.

The paper presents the implementation of two nonlinear noise reduction methods applied to speech enhancement. The methods are based on embedding the noisy signal in a high-dimensional reconstructed phase space and applying singular value decomposition to project the signal into a lower dimension. The advantages of these nonlinear methods include that they do not require explicit models of noise spectra and do not have the typical "musical tone" side effects associated with traditional linear speech enhancement methods. The proposed nonlinear methods are compared with traditional speech enhancement techniques, including spectral subtraction, Wiener filtering, and Ephraim-Malah filtering, on example speech utterances with additive white noise for a variety of SNR levels. The results show that the local nonlinear noise reduction method outperforms Wiener filtering and spectral subtraction, but not Ephraim-Malah filtering, as had been suggested by previous studies.

The paper presents a novel method for speech recognition by utilizing nonlinear/chaotic signal processing techniques to extract time-domain based phase space features. By exploiting the theoretical results derived in nonlinear dynamics, a processing space called a reconstructed phase space can be generated where a salient model (the natural distribution of the attractor) can be extracted for speech recognition. To discover the discriminatory power of these features, isolated phoneme classification experiments were performed using the TIMIT corpus and compared to a baseline classifier that uses MFCC (Mel frequency cepstral coefficient) features. The results demonstrate that phase space features contain substantial discriminatory power, even though MFCC features outperformed the phase space features on direct comparisons. The authors conjecture that phase space and MFCC features used in combination within a classifier may yield increased accuracy for various speech recognition tasks.

This paper presents a time-aligned singular value decomposition (SVD) analysis for speaker identification. SVD analysis has been used for fast spectral matching based on a global representation of an entire utterance. We incorporate temporal normalization directly into the decomposition by using a dynamic time warping (DTW) path to time-align the rows of the feature matrix prior to SVD analysis. Speaker identification results using the TI-46 database indicates that the time-aligned SVD significantly improves accuracy for most threshold choices.

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

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

Markov random fields (MRFs) continue to be an important and useful representation for modelling textured images. Standard methods for MRF image modeling make use of the equivalent Gibbs distribution (GD) to express the joint probabilities of groups of neighboring pixels. The authors investigate a new approach to the use of the GD in image modeling. Specifically, they develop an adaptive approach to the formation of clique potential functions for the distribution. Traditional tools, such as the multi-level logistic (MLL) model, have been based on the use of a predetermined and identical set of potential functions. In the present paper it is shown that by incorporating additional parameters into the model in order to control the shape of these functions, it is possible to arrive at a more complete parametrization of the image. A simple model based on this concept is described and implemented, and image simulations using the well-known Gibbs sampler algorithm are constructed to demonstrate the usefulness of an adaptive set of potential functions.>