Minimum generation error training for HMM-based prediction of articulatory movements

This paper describes an approach for improving a statistical parametric-based logF0 model using minimum-generationerror (MGE) training. Compared with the previous scheme based on decision tree clustering, MGE allows the minimisation of the error in the generated logF0 to take into account not only each cluster by itself, but also the way in which the clusters interact with each other in the generation of the F0 over the whole sentence. Moreover, the “weights” of each component of the model, which previously were adjusted manually, are optimized automatically by the MGE training during the re-estimation of the model covariances. Objective evaluation indicated that, although the logF0 contours generated by the models trained with MGE have approximately the same root mean square error and correlation factor as those generated with the baseline models, they present a higher dynamic range. The subjective evaluation shows a small but significant preference for the system trained with MGE.

In this paper, we propose a minimum generation error (MGE) training method to refine the audio-visual HMM to improve visual speech trajectory synthesis. Compared with the traditional maximum likelihood (ML) estimation, the proposed MGE training explicitly optimizes the quality of generated visual speech trajectory, where the audio-visual HMM modeling is jointly refined by using a heuristic method to find the optimal state alignment and a probabilistic descent algorithm to optimize the model parameters under the MGE criterion. In objective evaluation, compared with the ML-based method, the proposed MGE-based method achieves consistent improvement in the mean square error reduction, correlation increase, and recovery of global variance. It also improves the naturalness and audio-visual consistency perceptually in the subjective test.

Due to the inconsistency between the maximum likelihood (ML) based training and the synthesis application in HMM-based speech synthesis, a minimum generation error (MGE) criterion had been proposed for HMM training. This paper continues to apply the MGE criterion to model adaptation for HMM-based speech synthesis. We propose a MGE linear regression (MGELR) based model adaptation algorithm, where the regression matrices used to transform source models to target models are optimized to minimize the generation errors for the input speech data uttered by the target speaker. The proposed MGELR approach was compared with the maximum likelihood linear regression (MLLR) based model adaptation. Experimental results indicate that the generation errors were reduced after the MGELR-based model adaptation. And from the subjective listening test, the discrimination and the quality of the synthesized speech using MGELR were better than the results using MLLR.

Speech to Text was one of speech recognition applications which speech signal was processed, recognized and converted into a textual representation. Hidden Markov Model (HMM) was the widely used method in speech recognition. However, the level of accuracy using HMM was strongly influenced by the optimalization of extraction process and modellling methods. Hence in this research, the use of genetic algorithm (GA) method to optimize the Ergodic HMM was tested. In Hybrid HMM-GA, GA was used to optimize the Baum-welch method in the training process. It was useful to improve the accuracy of the recognition result which is produced by the HMM parameters that generate the low accuracy when the HMM are tested. Based on the research, the percentage increases the level of accuracy of 20% to 41%. Proved that the combination of GA in HMM method can gives more optimal results when compared with the HMM system that not combine with any method.

ABSTRACTUpdating of reservoir models by history matching of 4D seismic data along with production data gives us a better understanding of changes to the reservoir, reduces risk in forecasting and leads to better management decisions. This process of seismic history matching requires an accurate representation of predicted and observed data so that they can be compared quantitatively when using automated inversion. Observed seismic data is often obtained as a relative measure of the reservoir state or its change, however. The data, usually attribute maps, need to be calibrated to be compared to predictions. In this paper we describe an alternative approach where we normalize the data by scaling to the model data in regions where predictions are good. To remove measurements of high uncertainty and make normalization more effective, we use a measure of repeatability of the monitor surveys to filter the observed time‐lapse data.We apply this approach to the Nelson field. We normalize the 4D signature based on deriving a least squares regression equation between the observed and synthetic data which consist of attributes representing measured acoustic impedances and predictions from the model. Two regression equations are derived as part of the analysis. For one, the whole 4D signature map of the reservoir is used while in the second, 4D seismic data is used from the vicinity of wells with a good production match. The repeatability of time‐lapse seismic data is assessed using the normalized root mean square of measurements outside of the reservoir. Where normalized root mean square is high, observations and predictions are ignored. Net: gross and permeability are modified to improve the match.The best results are obtained by using the normalized root mean square filtered maps of the 4D signature which better constrain normalization. The misfit of the first six years of history data is reduced by 55 per cent while the forecast of the following three years is reduced by 29 per cent. The well based normalization uses fewer data when repeatability is used as a filter and the result is poorer. The value of seismic data is demonstrated from production matching only where the history and forecast misfit reductions are 45% and 20% respectively while the seismic misfit increases by 5%. In the best case using seismic data, it dropped by 6%. We conclude that normalization with repeatability based filtering is a useful approach in the absence of full calibration and improves the reliability of seismic data.

HighlightsSoftware was developed for estimation of DSSAT CSM-CROPGRO-Soybean cultivar coefficients.Phenology-related coefficients were estimated based on observed phenological events.Growth-related cultivar coefficients were estimated based on time-series observations.Cultivar coefficients were optimized based on single- and multiple-experiment data sets.Abstract. The Decision Support System for Agrotechnology Transfer (DSSAT) is one of the most popular software solutions for predicting crop growth and yield while capturing the effects of management practices and interactions between the crop and the environment. Accurate estimation of the crop cultivar coefficients that govern in-season growth and development is critical for correct yield estimates. The manual cultivar coefficient estimation process is time-consuming and results in user-dependent, subjective optimums that are difficult to reproduce. Typically, end-of-season observations (point-based) are used for estimating dynamic in-season biomass accumulation rates. The objective of this study was to develop a time-series estimator (TSE) capable of using multiple in-season observations for estimating the coefficients that define in-season growth and biomass partitioning. Using the TSE, cultivar coefficients were estimated based on multiple in-season observations of leaf area index (LAI) and shoot, leaf, and grain dry matter weights. The cultivar coefficients were estimated from single- and multiple-treatment (seasons and locations) in-season observations. This was done for two cultivars for six management × environment combinations. Estimated multiple-treatment based cultivar coefficients were evaluated with an independent data set and compared to DSSAT standard (manual) coefficients and the cultivar coefficients estimated with the GLUE method. The average normalized root mean squared error (nRSME) for LAI and shoot, leaf, and grain weights was 26% lower for one cultivar and about the same for the other cultivar when compared to the DSSAT standard. Because GLUE uses end-of-season point-based cultivar coefficient estimation, the grain weight over time was underestimated in earlier phases and more accurate toward harvest. The TSE-estimated cultivar coefficients based on 346 in-season observations across multiple target variables and six experiments more accurately reflected in-season growth and grain weight without compromising final grain weight predictions. Keywords: . CROPGRO-Soybean, DSSAT, Genetic coefficients, Normalized root mean square error minimization, Time-seris observations.

The study reported in this paper represents a continuation of our previous work on the application of Hidden Markov models (HMMs) to the translational and rotational invariant classification of SAR targets. The traditional method of making classification decisions using an HMM does not achieve the desired objective of minimizing the number of misclassifications. We present a novel technique that minimizes the probability of misclassification error. This approach, which is an adaptation of an existing Minimum Classification Error strategy, is globally optimal. The proposed method applies basic principles of pattern recognition to reduce the expected misclassification rate by dynamically perturbing the HMM parameters using a constraint on a cross-entropy measure and distance separation between pairs of HMM models. Like the traditional implementation of an HMM, our new formulation can still be implemented using an efficient forward-backward algorithm for estimating the HMM parameters. We tested our classifier on a public mixed- target MSTAR database and compared our approach to the original HMM approach trained by using a maximum likelihood criterion. The results indicate a significant improvement over the original HMM approach. Current scores from our method are about in excess of 90% on testing data sets.

In this paper, we propose a new objective evaluation method for hidden Markov model (HMM)-based speech synthesis using Kullback-Leibler divergence (KLD). The KLD is used to measure the difference between the probability density functions (PDFs) of the acoustic feature vectors extracted from natural training and synthetic speech data. For the evaluation, Gaussian mixture model (GMM) is used to model the distribution of acoustic feature vectors, including the fundamental frequency (F0). Continuous F0, obtained with linear interpolation, is used in the evaluation. In essence, the KLD is the expectation of the logarithmic difference between the likelihoods calculated on training and synthetic speech. This likelihood difference is appropriate to characterize the quality of a HMM-based speech synthesis system in generating synthetic speech using a maximum likelihood criterion. The objective evaluation is tested with 3 different HMM-based speech synthesis systems which use multi-space distribution (MSD) to model discontinuous F0. These systems are trained on a common speech corpus in French. We propose an index to evaluate HMM-based speech synthesis system which takes into account the relative variation of the KLDs on test sets of synthetic and natural speech. This index correlates inversely with the result of the MOS (mean opinion score) perceptual test.

Computational cancer phylogenetics can play an important role in delineating possible tumor progression pathways and identifying molecular subtypes and mechanisms of action. We previously developed a pipeline for constructing tumor phylogenies from recurring cell types computationally inferred from whole genome copy number data. The accuracy and detail of these tumor phylogenies, however, depends on the identification of accurate and high-resolution molecular markers of progression, i.e., reproducible regions of copy number variation that can be used to robustly differentiate different subtypes and stages of progression. Here we present a new method for the problem using hidden Markov models (HMMs) to derive robust, high resolution progression markers from sets of tumor samples. We demonstrate our method on a publicly available array comparative genome hybridization (aCGH) dataset (NCBI GEO GSE16672, Navin et al., 2010) from sectioned primary ductal breast tumors. Our method uses an HMM, a class of probabilistic models, to classify sets of aCGH data into a partitioning of samples into normal (diploid) or amplified at each copy number probe. It differs from other similar HMM methods primarily in seeking a parsimonious set of combinations of amplification states able to explain all aCGH profiles simultaneously in order to identify robust markers of progression across samples. The model learns frequencies with which different combinations of amplifications are observed across the samples by modeling individual probes as Gaussian random variables with either normal or tetraploid means, with data more consistent with tetraploid being classified as amplified and those more consistent with diploid classified as normal. To handle a combinatorial explosion in combinations of amplification states with increasing numbers of samples, the method introduces a Gibbs sampling algorithm to learn a parsimonious model of the most frequently occurring combinations of amplification states. We applied our methods to a previously constructed set of inferred cell types derived from the Navin et al. data (Tolliver et al., 2010) and to a comparison set of 9 random samples from the raw aCGH data. We validated our model relative to manual labeling of amplicons on the same data. In both experiments, the HMM method was able to pick up significantly larger numbers of robustly amplified segments per chromosome than did prior methods or manual analysis. The resulting segments can be directly fed into downstream analysis routines for phylogeny inference or other predictions. In future work, the HMM method may be improved by fine-tuning the underlying model for copy number variation. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 103rd Annual Meeting of the American Association for Cancer Research; 2012 Mar 31-Apr 4; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2012;72(8 Suppl):Abstract nr 3964. doi:1538-7445.AM2012-3964

Computational modeling, such as finite element analysis, is employed in a range of biomechanics specialties, including impact biomechanics and surgical planning. These models rely on accurate material properties for skeletal muscle, which comprises roughly 40% of the human body. Due to surrounding tissues, compressed skeletal muscle in vivo likely experiences a semi-confined state. Nearly all previous studies investigating passively compressed muscle at the tissue level have focused on muscle in unconfined compression. The goals of this study were to (1) examine the stiffness and time-dependent material properties of skeletal muscle subjected to both confined and unconfined compression (2) develop a model that captures passive muscle mechanics under both conditions and (3) determine the extent to which different assumptions of volumetric behavior affect model results. Muscle in confined compression exhibited stiffer behavior, agreeing with previous assumptions of near-incompressibility. Stress relaxation was found to be faster under unconfined compression, suggesting there may be different mechanisms that support load these two conditions. Finite element calibration was achieved through nonlinear optimization (normalized root mean square error <6%) and model validation was strong (normalized root mean square error <17%). Comparisons to commonly employed assumptions of bulk behavior showed that a simple one parameter approach does not accurately simulate confined compression. We thus recommend the use of a properly calibrated, nonlinear bulk constitutive model for modeling of skeletal muscle in vivo. Future work to determine mechanisms of passive muscle stiffness would enhance the efforts presented here.

The first lactation weekly test day milk yield (WTDMY) records (57233) and monthly test day milk yield (MTDMY) records (11970) spread over a period of 23 years (1984– 2006) sired by 82 bulls of Karan Fries cattle were used to develop the best lactation curve model. The lactation curve parameters of Quadratic cum log model (QCLM), Gamma function (GF), Cobby Le Du model (CLDM), Polynomial regression function (PRF) and Multiphasic logistic function (MLF) were estimated. The average weekly test day milk yield was predicted with high degree of accuracy (R2 &gt; 85%) by all the models with the maximum accuracy (R2 = 99.50%) obtained by polynomial regression function (PRF) and the least fit (R2 = 87.90%) was obtained by Gamma function (GF). However Quadratic cum log model (R2 = 99.20%) was almost equal to polynomial regression function and was better than Cobby Le Du model (R2 = 92.80%). The average root mean square error (RMSE) was found to be minimum with PRF (0.0121 Kg) followed by MLF (Triphasic). Thus the best fit model was polynomial regression function, which was better than other functions for prediction of first lactation WTDMY. The average monthly test day milk yield was predicted with high degree of accuracy (R2 &gt; 75%) by all the models with the maximum accuracy (R2 &gt; 99.41%) obtained by polynomial regression function (PRF) and the least fit was obtained with gamma function (R2 = 79.05%). However, quadratic-cum-log model (R2 = 99.28%) was almost equal to polynomial regression function and was better than CLDM (R2 = 89.50%). The average root mean square error (RMSE) was found to be minimum with PRF (0.0061 kg) followed by QCLM (0.0620). Thus the best fit model was polynomial regression function for prediction of both weekly and monthly test day milk yield of first lactation in Karan Fries cows.

Phenological shifts in plant species are one of the most conspicuous signs of climate change impact on the biosphere. Modeling phenological variations of plant species over broad regions is challenging because of the varied climatic requirements of geographic populations due to local adaptation. In this study, we developed an empirical method to calibrate phenological models of temperate trees using latitude as a predictor to account for local adaptation of populations to a N-S temperature gradient. Fourteen widely distributed tree species in the eastern U.S.A. were investigated using data from the USA-National Phenology Network. We implemented the method in a basic thermal time bud break model to introduce the algorithm of the method and test its effectiveness. For each species, dates of breaking leaf buds were first predicted using a traditional non-spatial model and then with a spatial model that has the critical thermal forcing requirements calibrated for different populations at varied latitudes. As anticipated, non-spatial model predictions that assumed a uniform forcing requirement across latitudes showed consistent and systematic biases at both higher (overestimation-predictions being later) and lower (underestimation-predictions being earlier) latitudes. Spatial models that have been calibrated using our method removed the geographic biases and yielded latitudinal gradients that more closely matched those of the observations. The spatial models also reduced the overall prediction errors from an average root mean square error (RMSE) of 32.2 days to 20.4 days for the training dataset and an average root mean square error for prediction (RMSEP) of 32.2 days to 19.9 days for the testing dataset. This paper is focused on introducing the new calibration method as a preparatory step toward developing operational models that may potentially predict large-scale and range-wide phenological responses of various plant species to climatic changes with improved local accuracy.

Background/Purpose: The acquisition of three-dimensional ground reaction forces (3D-GRF) during running is crucial for monitoring and assessing running performance. However, limitations often arise in collecting 3D-GRF due to equipment and environmental constraints. Inertial sensors are gaining popularity in running applications because of their convenience, broad applicability, and cost-effectiveness. This study aimed to utilize an inertial sensor for gathering kinematic data during running, establish a high-precision neural network model for continuous 3D-GRF estimation, and investigate estimation variances across different body segments. Method: A total of thirty-three college students (17 male, 16 female) participated in this study. Kinematic and kinetic data were concurrently recorded using inertial sensors placed on the lower back, lateral thigh, anterior lower leg, and dorsum of the foot, while three-dimensional force plates were utilized during running at speeds of 8/10/12 km/h. The dataset was partitioned into training, validation, and test sets at a ratio of 24:3:6. A Long Short-Term Memory Neural Network (LSTM) was developed for training to predict 3D-GRF. Evaluation metrics included coefficients of determination, Root Mean Square Error (RMSE), and Normalized Root Mean Square Error (NRMSE). Models customized for specific body segments underwent a nine-fold cross-validation procedure, employing one-way ANOVA to evaluate estimation variances across different anatomical sites. Results: Estimation outcomes at different sites indicated optimal estimation accuracy at the right thigh. The coefficients of determination for Anteroposterior (AP)-GRF, Mediolateral (ML)-GRF, and Vertical (V)-GRF were 0.81±0.00, 0.55±0.14, and 0.97±0.02, respectively. RMSE values were 0.06±0.01BW, 0.06±0.01BW, and 0.14±0.03BW. Significant variances were observed between the right thigh and other sites in the horizontal plane (p &lt; 0.05). The results from the test set showed correlation coefficients, RMSE, and NRMSE for AP-GRF, ML-GRF, and V-GRF: AP-GRF, 0.94±0.01, 0.06±0.01BW, 4.78%±0.50%; ML-GRF, 0.83±0.08, 0.10±0.01BW, 9.70%±2.41%; V-GRF, 0.99±0.00, 0.14±0.02BW, 2.79%±0.24%. Conclusion/Discussion: 1. The neural network model developed in this study demonstrated exceptional accuracy across three dimensions, facilitating continuous estimation of 3D ground reaction forces during running. Correlation coefficients surpassed 0.8 in every direction, with RMSE values below 0.2BW and NRMSE below 5%. 2. Estimation results from inertial sensors positioned at different sites exhibited diverse levels of disparity. The lateral thigh emerged as the most proficient location for estimation, with the lower back excelling in vertical ground reaction force estimation but displaying comparatively weaker performance in horizontal ground reaction force estimation.

A robust model for diagnosing water stress of winter wheat by combining UAV multispectral and thermal remote sensing

There is uncertainty in the neuromusculoskeletal system, and deterministic models cannot describe this significant presence of uncertainty, affecting the accuracy of model predictions. In this paper, a knee joint angle prediction model based on surface electromyography (sEMG) signals is proposed. To address the instability of EMG signals and the uncertainty of the neuromusculoskeletal system, a non-parametric probabilistic model is developed using a Gaussian process model combined with the physiological properties of muscle activation. Since the neuromusculoskeletal system is a dynamic system, the Gaussian process model is further combined with a non-linear autoregressive with eXogenous inputs (NARX) model to create a Gaussian process autoregression model. In this paper, the normalized root mean square error (NRMSE) and the correlation coefficient (CC) are compared between the joint angle prediction results of the Gaussian process autoregressive model prediction and the actual joint angle under three test scenarios: speed-dependent, multi-speed and speed-independent. The mean of NRMSE and the mean of CC for all test scenarios in the healthy subjects dataset and the hemiplegic patients dataset outperform the results of the Gaussian process model, with significant differences (p < 0.05 and p < 0.05, p < 0.05 and p < 0.05). From the perspective of uncertainty, a non-parametric probabilistic model for joint angle prediction is established by using Gaussian process autoregressive model to achieve accurate prediction of human movement.