Sparse Training Data Research Articles

Eigenvoice modelling for cross likelihood ratio based speaker clustering: A Bayesian approach

This paper proposes the use of Bayesian approaches with the cross likelihood ratio (CLR) as a criterion for speaker clustering within a speaker diarization system, using eigenvoice modelling techniques. The CLR has previously been shown to be an effective decision criterion for speaker clustering using Gaussian mixture models. Recently, eigenvoice modelling has become an increasingly popular technique, due to its ability to adequately represent a speaker based on sparse training data, as well as to provide an improved capture of differences in speaker characteristics. The integration of eigenvoice modelling into the CLR framework to capitalize on the advantage of both techniques has also been shown to be beneficial for the speaker clustering task. Building on that success, this paper proposes the use of Bayesian methods to compute the conditional probabilities in computing the CLR, thus effectively combining the eigenvoice-CLR framework with the advantages of a Bayesian approach to the diarization problem. Results obtained on the 2002 Rich Transcription (RT-02) Evaluation dataset show an improved clustering performance, resulting in a 33.5% relative improvement in the overall diarization error rate (DER) compared to the baseline system.

Mixture of Factor Analyzers Using Priors From Non-Parallel Speech for Voice Conversion

A robust voice conversion function relies on a large amount of parallel training data, which is difficult to collect in practice. To tackle the sparse parallel training data problem in voice conversion, this paper describes a mixture of factor analyzers method which integrates prior knowledge from non-parallel speech into the training of conversion function. The experiments on CMU ARCTIC corpus show that the proposed method improves the quality and similarity of converted speech. With both objective and subjective evaluations, we show the proposed method outperforms the baseline GMM method.

Detail-Preserving Controllable Deformation from Sparse Examples

Recent advances in laser scanning technology have made it possible to faithfully scan a real object with tiny geometric details, such as pores and wrinkles. However, a faithful digital model should not only capture static details of the real counterpart but also be able to reproduce the deformed versions of such details. In this paper, we develop a data-driven model that has two components; the first accommodates smooth large-scale deformations and the second captures high-resolution details. Large-scale deformations are based on a nonlinear mapping between sparse control points and bone transformations. A global mapping, however, would fail to synthesize realistic geometries from sparse examples, for highly deformable models with a large range of motion. The key is to train a collection of mappings defined over regions locally in both the geometry and the pose space. Deformable fine-scale details are generated from a second nonlinear mapping between the control points and per-vertex displacements. We apply our modeling scheme to scanned human hand models, scanned face models, face models reconstructed from multiview video sequences, and manually constructed dinosaur models. Experiments show that our deformation models, learned from extremely sparse training data, are effective and robust in synthesizing highly deformable models with rich fine features, for keyframe animation as well as performance-driven animation. We also compare our results with those obtained by alternative techniques.

Optimal Viterbi Bayesian predictive classification for data from finite alphabets

A family of Viterbi Bayesian predictive classifiers has been recently popularized for speech recognition applications with continuous acoustic signals modeled by finite mixture densities embedded in a hidden Markov framework. Here we generalize such classifiers to sequentially observed data from multiple finite alphabets and derive the optimal predictive classifier under exchangeability of the emitted symbols. We demonstrate that the optimal predictive classifier which learns from unlabelled test items improves considerably upon marginal maximum a posteriori rule in the presence of sparse training data. It is shown that the learning process saturates when the amount of test data tends to infinity, such that no further gain in classification accuracy is possible upon arrival of new test items in the long run.

A Cluster Adaptive Training Algorithm for Text-Independent Speaker Verification with Sparse Training Data

A Cluster Adaptive Training Algorithm for Text-Independent Speaker Verification with Sparse Training Data

Automatic categorization for improving Spanish into Spanish Sign Language machine translation

This paper describes a preprocessing module for improving the performance of a Spanish into Spanish Sign Language (Lengua de Signos Española: LSE) translation system when dealing with sparse training data. This preprocessing module replaces Spanish words with associated tags. The list with Spanish words (vocabulary) and associated tags used by this module is computed automatically considering those signs that show the highest probability of being the translation of every Spanish word. This automatic tag extraction has been compared to a manual strategy achieving almost the same improvement. In this analysis, several alternatives for dealing with non-relevant words have been studied. Non-relevant words are Spanish words not assigned to any sign. The preprocessing module has been incorporated into two well-known statistical translation architectures: a phrase-based system and a Statistical Finite State Transducer (SFST). This system has been developed for a specific application domain: the renewal of Identity Documents and Driver's License. In order to evaluate the system a parallel corpus made up of 4080 Spanish sentences and their LSE translation has been used. The evaluation results revealed a significant performance improvement when including this preprocessing module. In the phrase-based system, the proposed module has given rise to an increase in BLEU (Bilingual Evaluation Understudy) from 73.8% to 81.0% and an increase in the human evaluation score from 0.64 to 0.83. In the case of SFST, BLEU increased from 70.6% to 78.4% and the human evaluation score from 0.65 to 0.82.

Syntax-based reordering for statistical machine translation

Abstract In this paper, we develop an approach called syntax-based reordering (SBR) to handling the fundamental problem of word ordering for statistical machine translation (SMT). We propose to alleviate the word order challenge including morpho-syntactical and statistical information in the context of a pre-translation reordering framework aimed at capturing short- and long-distance word distortion dependencies. We examine the proposed approach from the theoretical and experimental points of view discussing and analyzing its advantages and limitations in comparison with some of the state-of-the-art reordering methods. In the final part of the paper, we describe the results of applying the syntax-based model to translation tasks with a great need for reordering (Chinese-to-English and Arabic-to-English). The experiments are carried out on standard phrase-based and alternative N-gram-based SMT systems. We first investigate sparse training data scenarios, in which the translation and reordering models are trained on a sparse bilingual data, then scaling the method to a large training set and demonstrating that the improvement in terms of translation quality is maintained.

Manifold‐Based 3D Face Caricature Generation with Individualized Facial Feature Extraction

AbstractCaricature is an interesting art to express exaggerated views of different persons and things through drawing. The face caricature is popular and widely used for different applications. To do this, we have to properly extract unique/specialized features of a person's face. A person's facial feature not only depends on his/her natural appearance, but also the associated expression style. Therefore, we would like to extract the neutural facial features and personal expression style for different applicaions. In this paper, we represent the 3D neutral face models in BU–3DFE database by sparse signal decomposition in the training phase. With this decomposition, the sparse training data can be used for robust linear subspace modeling of public faces. For an input 3D face model, we fit the model and decompose the 3D model geometry into a neutral face and the expression deformation separately. The neutral geomertry can be further decomposed into public face and individualized facial feature. We exaggerate the facial features and the expressions by estimating the probability on the corresponding manifold. The public face, the exaggerated facial features and the exaggerated expression are combined to synthesize a 3D caricature for a 3D face model. The proposed algorithm is automatic and can effectively extract the individualized facial features from an input 3D face model to create 3D face caricature.

A segment selection technique for speaker verification

The performance of speaker verification systems degrades considerably when the test segments are utterances of very short duration. This might be either due to variations in score-matching arising from the unobserved speech sounds of short speech utterances or the fact that the shorter the utterance, the greater the effect of individual speech sounds on the average likelihood score. In other words, the effects of individual speech sounds have not been cancelled out by a large number of speech sounds in very short utterances. This paper presents a score-based segment selection technique for discarding portions of speech that result in poor discrimination ability in a speaker verification task. Theory is developed to detect the most significant and reliable speech segments based on the probability that the test segment comes from a fixed set of cohort models. This approach, suitable for any duration of test utterance, reduces the effect of acoustic regions of the speech that are not accurately modelled due to sparse training data, and makes a decision based only on the segments that provide the best-matched scores from the segment selection algorithm. The proposed segment selection technique provides reductions in relative error rate of 22% and 7% in terms of minimum Detection Cost Function (DCF) and Equal Error Rate (EER) compared with a baseline used the segment-based normalization, when evaluated on the short utterances of NIST 2002 Speaker Recognition Evaluation dataset.

Utilizing hierarchical feature domain values for prediction

We propose a Bayesian learning framework which can exploit hierarchical structures of discrete feature domain values to improve the prediction performance on sparse training data. One characteristic of our framework is that it provides a principled way based on mean–variance analysis to transform an original feature domain value to a coarser granularity by exploiting the underlying hierarchical structure. Through this transformation, a tradeoff between precision and robustness is achieved to improve the parameter estimation for prediction. We have conducted comparative experiments using three real-world data sets. The results demonstrate that utilizing domain value hierarchies gains benefits for prediction.

Achieving a reliable compact acoustic model for embedded speech recognition system with high confusion frequency model handling

An acoustic model for an embedded speech recognition system must exhibit two desirable features; the ability to minimize the performance degradation in recognition, while solving the memory problem under the constraint of limited system resources. Moreover, for general speech recognition tasks, context dependent models such as state-clustered tri-phones are used to guarantee the high recognition performance of the embedded system. To cope with these challenges, we introduce the state-clustered tied-mixture (SCTM) HMM as a method of optimizing an acoustic model. The proposed SCTM modeling system offers a significant improvement in recognition performance, as well as providing a solution to sparse training data problems. Moreover, the state weight quantizing method achieves a drastic reduction in the size of the model. However, using models constructed only in this way is insufficient to improve the recognition rate in some tasks where a large mutual similarity exists, such as in the case of the Korean-digit recognition task. Hence, we also construct new dedicated HMM’s for all or part of the Korean-digits that have exclusive states using the same Gaussian pool of previous tri-phone models. In this paper, we describe the acoustic model optimization procedure for embedded speech recognition systems and the corresponding performance evaluation results.

Eigenvoice modeling with sparse training data

We derive an exact solution to the problem of maximum likelihood estimation of the supervector covariance matrix used in extended MAP (or EMAP) speaker adaptation and show how it can be regarded as a new method of eigenvoice estimation. Unlike other approaches to the problem of estimating eigenvoices in situations where speaker-dependent training is not feasible, our method enables us to estimate as many eigenvoices from a given training set as there are training speakers. In the limit as the amount of training data for each speaker tends to infinity, it is equivalent to cluster adaptive training.

Parameter Sharing in Mixture of Factor Analyzers for Speaker Identification

This paper investigates the parameter tying structures of a mixture of factor analyzers (MFA) and discriminative training of MFA for speaker identification. The parameters of factor loading matrices or diagonal matrices are shared in different mixtures of MFA. Then, minimum classification error (MCE) training is applied to the MFA parameters to enhance the discrimination ability. The result of a text-independent speaker identification experiment shows that MFA outperforms the conventional Gaussian mixture model (GMM) with diagonal or full covariance matrices and achieves the best performance when sharing the diagonal matrices, resulting in a relative gain of 26% over the GMM with diagonal covariance matrices. The improvement is more significant especially in sparse training data condition. The recognition performance is further improved by MCE training with an additional gain of 3% error reduction.

Similarity searching of chemical databases using atom environment descriptors (MOLPRINT 2D): evaluation of performance.

A molecular similarity searching technique based on atom environments, information-gain-based feature selection, and the naive Bayesian classifier has been applied to a series of diverse datasets and its performance compared to those of alternative searching methods. Atom environments are count vectors of heavy atoms present at a topological distance from each heavy atom of a molecular structure. In this application, using a recently published dataset of more than 100000 molecules from the MDL Drug Data Report database, the atom environment approach appears to outperform fusion of ranking scores as well as binary kernel discrimination, which are both used in combination with Unity fingerprints. Overall retrieval rates among the top 5% of the sorted library are nearly 10% better (more than 14% better in relative numbers) than those of the second best method, Unity fingerprints and binary kernel discrimination. In 10 out of 11 sets of active compounds the combination of atom environments and the naive Bayesian classifier appears to be the superior method, while in the remaining dataset, data fusion and binary kernel discrimination in combination with Unity fingerprints is the method of choice. Binary kernel discrimination in combination with Unity fingerprints generally comes second in performance overall. The difference in performance can largely be attributed to the different molecular descriptors used. Atom environments outperform Unity fingerprints by a large margin if the combination of these descriptors with the Tanimoto coefficient is compared. The naive Bayesian classifier in combination with information-gain-based feature selection and selection of a sensible number of features performs about as well as binary kernel discrimination in experiments where these classification methods are compared. When used on a monoaminooxidase dataset, atom environments and the naive Bayesian classifier perform as well as binary kernel discrimination in the case of a 50/50 split of training and test compounds. In the case of sparse training data, binary kernel discrimination is found to be superior on this particular dataset. On a third dataset, the atom environment descriptor shows higher retrieval rates than other 2D fingerprints tested here when used in combination with the Tanimoto similarity coefficient. Feature selection is shown to be a crucial step in determining the performance of the algorithm. The representation of molecules by atom environments is found to be more effective than Unity fingerprints for the type of biological receptor similarity calculations examined here. Combining information prior to scoring and including information about inactive compounds, as in the Bayesian classifier and binary kernel discrimination, is found to be superior to posterior data fusion (in the datasets tested here).

A method for dynamics identification for haptic display of the operating feel in virtual environments

Realistic dynamics models are important for haptic display for virtual reality systems. Such dynamic models are desirably obtained via experimental identification. However, traditional dynamics identification methods normally require large sized training data sets, which maybe difficult to meet in many practical applications. To obtain the dynamics models, we present, in this paper, an identification method using support vector machines regression algorithm which is more effective than traditional methods for sparse training data. This method can be used as a generic learning machine or as a special learning technique that can make full use of the available knowledge about the dynamics structure. The experimental results show the application of our method for identifying friction models for haptic display.

Multigrained modeling with pattern specific maximum likelihood transformations for text-independent speaker recognition

We present a transformation-based, multigrained data modeling technique in the context of text independent speaker recognition, aimed at mitigating difficulties caused by sparse training and test data. Both identification and verification are addressed, where we view the entire population as divided into the target population and its complement, which we refer to as the background population. First, we present our development of maximum likelihood transformation based recognition with diagonally constrained Gaussian mixture models and show its robustness to data scarcity with results on identification. Then for each target and background speaker, a multigrained model is constructed using the transformation based extension as a building block. The training data is labeled with an HMM based phone labeler. We then make use of a graduated phone class structure to train the speaker model at various levels of detail. This structure is a tree with the root node containing all the phones. Subsequent levels partition the phones into increasingly finer grained linguistic classes. This method affords the use of fine detail where possible, i.e., as reflected in the amount of training data distributed to each tree node. We demonstrate the effectiveness of the modeling with verification experiments in matched and mismatched conditions.

On the limitations of neural network techniques for analysing cause and effect relationships in manufacturing processes - A case study

The cause and effect relationship is complex for many manufacturing processes. The ability to learn causal relationships from diagnostic examples is extremely useful. This learning ability can help not only to quantify the influence of causes on defects for existing components but also to set up new process, material and design parameters to manufacture new components. A neural network approach that can adapt and learn from past examples, was explored for analysing and quantifying cause and effect relationships. Neural networks use the data to extract any pre-existing relationships between the input and output variables. However, in many real situations very few good quality training examples are generally available. A few limitations of multi layer feed forward neural networks in presence of such noisy, limited and sparse training data sets have been discovered, and because of such limitations it has been found difficult to use this technique as a robust tool by end users in a foundry environment or any other manufacturing industry to analyse cause and effect relationships.

Comparison of fault detection techniques

Abstract Haar Wave-Net (HWN) and Projection Pursuit Regression (PPR) are two useful modeling tools for pattern classification. In this study, the two methodologies are compared with respect to the problem of misclassification close to class boundaries with sparse training data. A variety of examples were specifically tailored to elucidate their respective properties. It is observed that PPR locates the class boundaries at the midline of two classes of training data, which is a logical choice for the class boundary location, in the absence of sufficient information. For HWN, both the initial positioning of receptive fields and the density of training data near the class boundary may have great impact on the definition of the class boundary. Additionally, PPR and HWN are also compared to the Backpropagation Network (BPN), a standard technique for fault detection, with respect to their sensitivity to noise. The orthonormal and localized properties of the Haar basis functions enable a HWN to limit the noise effect within its local receptive fields. BPN propagates the noise effect throughout the input space. PPR provides a good tradeoff between reasonable generalization and noise localization. The fault diagnosis problem is investigated in a CSTR process, at both steady state and dynamic conditions. It is found that, for the dynamic case, the misclassification close to the class boundary is often due to lack of system observability.

Hybrid fuzzy modeling of chemical processes

Fuzzy models have been proved to have the ability of modeling all plants without any priori information. However, the performance of conventional fuzzy models can be very poor in the case of insufficient training data due to their poor extrapolation capacity. In order to overcome this problem, a hybrid grey-box fuzzy modeling approach is proposed in this paper to combine expert experience, local linear models and historical data into a uniform framework. It consists of two layers. The expert fuzzy model constructed from linguistic information, the local linear model and the T-S type fuzzy model constructed from data are all put in the first layer. Layer 2 is a fuzzy decision module that is used to decide which model in the first layer should be employed to make the final prediction. The output of the second layer is~the output of the hybrid fuzzy model. With the help of the linguistic information, the poor extrapolation capacity problem caused by sparse training data for conventional fuzzy models can be overcome. Simulation result for pH neutralization process demonstrates its modeling ability over the linear models, the expert fuzzy model and the conventional fuzzy model.

On adaptive decision rules and decision parameter adaptation for automatic speech recognition

Recent advances in automatic speech recognition are accomplished by designing a plug-in maximum a posteriori decision rule such that the forms of the acoustic and language model distributions are specified and the parameters of the assumed distributions are estimated from a collection of speech and language training corpora. Maximum-likelihood point estimation is by far the most prevailing training method. However, due to the problems of unknown speech distributions, sparse training data, high spectral and temporal variabilities in speech, and possible mismatch between training and testing conditions, a dynamic training strategy is needed. To cope with the changing speakers and speaking conditions in real operational conditions for high-performance speech recognition, such paradigms incorporate a small amount of speaker and environment specific adaptation data into the training process. Bayesian adaptive learning is an optimal way to combine prior knowledge in an existing collection of general models with a new set of condition-specific adaptation data. In this paper, the mathematical framework for Bayesian adaptation of acoustic and language model parameters is first described. Maximum a posteriori point estimation is then developed for hidden Markov models and a number of useful parameters densities commonly used in automatic speech recognition and natural language processing.