A Comparative Analysis of Bayesian Nonparametric Inference Algorithms for Acoustic Modeling in Speech Recognition

This paper describes the combination of two recent machine learning techniques for acoustic modeling in speech recognition: deep neural networks (DNNs) and graph-based semi-supervised learning (SSL). While DNNs have been shown to be powerful supervised classifiers and have achieved considerable success in speech recognition, graph-based SSL can exploit valuable complementary information derived from the manifold structure of the unlabeled test data. Previous work on graph-based SSL in acoustic modeling has been limited to frame-level classification tasks and has not been compared to, or integrated with, state-of-the-art DNN/HMM recognizers. This paper represents the first integration of graph-based SSL with DNN based speech recognition and analyzes its effect on word recognition performance. The approach is evaluated on two small vocabulary speech recognition tasks and shows a significant improvement in HMM state classification accuracy as well as a consistent reduction in word error rate over a state-of-the-art DNN/HMM baseline.

This paper describes a data-driven technique for optimizing the acoustic models for speech recognition systems that target commercial applications over telephones. Frame-averaged foreground log-likelihoods (foreground scores) correlate to recognition errors. These scores are used together with gender to optimize data weighting for the acoustic model. This process is interpreted as increasing the priors and associated parameters for poorly modeled data. The score-based optimization leads to about 7% fewer semantic errors on a live evaluation set collected after the last data used to estimate the acoustic model.

Hidden Markov Random Field (HMRF) models are important statistical tools for image segmentation. However, such models cannot automatically determine the number of clusters. In many practical applications, to determine the number of clusters in advance is impossible. The Dirichlet process mixture (DPM) models are nonparametric Bayesian models which can automatically determine the number of clusters for the clustering problem. A new HMRF model which combines the DPM model with Student's t-distribution is proposed in this paper. Compared to the Gaussian distribution, the Student's t-distribution with heavier-tailed can have more applications. This model can automatically determine the appropriate clustering number from the observed data. In addition, an efficient variational Bayesian inference algorithm is derived to the proposed model. Experimental results show that the proposed model outperforms other DPM-based models in image segmentation.

Modeling heterogeneous data sources remains a fundamental challenge of acoustic modeling in speech recognition. We call this the multi-condition problem because the speech data come from many different conditions. In this paper, we introduce the fundamental confusability problem in multi-condition learning, then discuss the problem formalization, the taxonomy, and the architectures for multi-condition learning. While the ideas presented are applicable to all classifiers, we focus our attention in this work on acoustic models based on deep neural networks (DNN). We propose four different strategies for multi-condition learning of a DNN that we refer to as a mixed-condition model, a condition-dependent model, a condition-normalizing model, and a condition-aware model. Based on the experimental results on the voice search and short message dictation task and the Aurora 4 task, we show that the confusability introduced when modeling heterogeneous data depends on the source of acoustic distortion itself, the front-end feature extractor, and the classifier. We also demonstrate the best approach for dealing with heterogeneous data may not be to let the model sort it out blindly, even with a classifier as sophisticated as a DNN.

Despite their known weaknesses, hidden Markov models (HMMs) have been the dominant technique for acoustic modeling in speech recognition for over two decades. Still, the advances in the HMM framework have not solved its key problems: it discards information about time dependencies and is prone to overgeneralization. In this paper, we attempt to overcome these problems by relying on straightforward template matching. The basis for the recognizer is the well-known DTW algorithm. However, classical DTW continuous speech recognition results in an explosion of the search space. The traditional top-down search is therefore complemented with a data-driven selection of candidates for DTW alignment. We also extend the DTW framework with a flexible subword unit mechanism and a class sensitive distance measure-two components suggested by state-of-the-art HMM systems. The added flexibility of the unit selection in the template-based framework leads to new approaches to speaker and environment adaptation. The template matching system reaches a performance somewhat worse than the best published HMM results for the Resource Management benchmark, but thanks to complementarity of errors between the HMM and DTW systems, the combination of both leads to a decrease in word error rate with 17% compared to the HMM results

We study large-scale kernel methods for acoustic modeling in speech recognition and compare their performance to deep neural networks (DNNs). We perform experiments on four speech recognition datasets, including the TIMIT and Broadcast News benchmark tasks, and compare these two types of models on frame-level performance metrics (accuracy, cross-entropy), as well as on recognition metrics (word/character error rate). In order to scale kernel methods to these large datasets, we use the random Fourier feature method of Rahimi and Recht [2007]. We propose two novel techniques for improving the performance of kernel acoustic models. First, in order to reduce the number of random features required by kernel models, we propose a simple but effective method for feature selection. The method is able to explore a large number of non-linear features while maintaining a compact model more efficiently than existing approaches. Second, we present a number of frame-level metrics which correlate very strongly with recognition performance when computed on the heldout set; we take advantage of these correlations by monitoring these metrics during training in order to decide when to stop learning. This technique can noticeably improve the recognition performance of both DNN and kernel models, while narrowing the gap between them. Additionally, we show that the linear bottleneck method of Sainath et al. [2013a] improves the performance of our kernel models significantly, in addition to speeding up training and making the models more compact. Together, these three methods dramatically improve the performance of kernel acoustic models, making their performance comparable to DNNs on the tasks we explored.

Automatic recognition of Arabic dialectal speech is a challenging task because Arabic dialects are essentially spoken varieties. Only few dialectal resources are available to date; moreover, most available acoustic data collections are transcribed without diacritics. Such a transcription omits essential pronunciation information about a word, such as short vowels. In this paper we investigate various procedures that enable us to use such training data by automatically inserting the missing diacritics into the transcription. These procedures use acoustic information in combination with different levels of morphological and contextual constraints. We evaluate their performance against manually diacritized transcriptions. In addition, we demonstrate the effect of their accuracy on the recognition performance of acoustic models trained on automatically diacritized training data.

Hidden Markov random field (HMRF) models are widely used for image segmentation, as they appear naturally in problems where a spatially constrained clustering scheme is asked for. A major limitation of HMRF models concerns the automatic selection of the proper number of their states, i.e., the number of region clusters derived by the image segmentation procedure. Existing methods, including likelihood- or entropy-based criteria, and reversible Markov chain Monte Carlo methods, usually tend to yield noisy model size estimates while imposing heavy computational requirements. Recently, Dirichlet process (DP, infinite) mixture models have emerged in the cornerstone of nonparametric Bayesian statistics as promising candidates for clustering applications where the number of clusters is unknown a priori; infinite mixture models based on the original DP or spatially constrained variants of it have been applied in unsupervised image segmentation applications showing promising results. Under this motivation, to resolve the aforementioned issues of HMRF models, in this paper, we introduce a nonparametric Bayesian formulation for the HMRF model, the infinite HMRF model, formulated on the basis of a joint Dirichlet process mixture (DPM) and Markov random field (MRF) construction. We derive an efficient variational Bayesian inference algorithm for the proposed model, and we experimentally demonstrate its advantages over competing methodologies.

This dissertation focuses on the development of multiple imputation models and algorithms for high-dimensional data with variable selection structures. Leveraging on the multivariate linear mixed-effects model with missing responses for clustered data, we incorporate the variable selection routines using spike-and-slab priors within the Bayesian variable selection framework. Specific choice of these priors allow us to "force'' variables of importance (e.g. design variables or variables known to play role in missingness mechanism) into the imputation models. Our ultimate goal is to improve computational speed by removing unnecessary variables. Markov chain Monte Carlo techniques have been designed to sample from the implied posterior distributions for model unknowns as well as missing data. A computationally efficient alternative, namely, variational inference algorithms have also been developed to overcome the computational burden of Markov chain Monte Carlo for sampling the posterior predictive distribution missing data. The continuous imputation model and algorithm can be easily modified through a calibration-based rounding strategy for multiple imputation of categorical variables. We also compare two imputation strategies, i.e., the sequential imputation and the joint modeling imputation, has been made in the presence of variable selection structures. We design a joint modeling based imputation framework using a nonparametric Bayesian Dirichlet process mixture model together with a sparse precision matrix model. Two computational algorithms have been carefully designed: the Gibbs sampler and the variational Bayesian inference algorithm. The proposed methodology can be applied to both continuous and mixed continuous and categorical missing data.

The aim of this study was to implement a Dirichlet process mixture (DPM) model for automatic tumor edge identification on (18)F-fluorodeoxyglucose positron emission tomography ((18)F-FDG PET) images by optimizing the parameters on which the algorithm depends, to validate it experimentally, and to test its robustness. The DPM model belongs to the class of the Bayesian nonparametric models and uses the Dirichlet process prior for flexible nonparametric mixture modeling, without any preliminary choice of the number of mixture components. The DPM algorithm implemented in the statistical software package R was used in this work. The contouring accuracy was evaluated on several image data sets: on an IEC phantom (spherical inserts with diameter in the range 10-37 mm) acquired by a Philips Gemini Big Bore PET-CT scanner, using 9 different target-to-background ratios (TBRs) from 2.5 to 70; on a digital phantom simulating spherical/uniform lesions and tumors, irregular in shape and activity; and on 20 clinical cases (10 lung and 10 esophageal cancer patients). The influence of the DPM parameters on contour generation was studied in two steps. In the first one, only the IEC spheres having diameters of 22 and 37 mm and a sphere of the digital phantom (41.6 mm diameter) were studied by varying the main parameters until the diameter of the spheres was obtained within 0.2% of the true value. In the second step, the results obtained for this training set were applied to the entire data set to determine DPM based volumes of all available lesions. These volumes were compared to those obtained by applying already known algorithms (Gaussian mixture model and gradient-based) and to true values, when available. Only one parameter was found able to significantly influence segmentation accuracy (ANOVA test). This parameter was linearly connected to the uptake variance of the tested region of interest (ROI). In the first step of the study, a calibration curve was determined to automatically generate the optimal parameter from the variance of the ROI. This "calibration curve" was then applied to contour the whole data set. The accuracy (mean discrepancy between DPM model-based contours and reference contours) of volume estimation was below (1 ± 7)% on the whole data set (1 SD). The overlap between true and automatically segmented contours, measured by the Dice similarity coefficient, was 0.93 with a SD of 0.03. The proposed DPM model was able to accurately reproduce known volumes of FDG concentration, with high overlap between segmented and true volumes. For all the analyzed inserts of the IEC phantom, the algorithm proved to be robust to variations in radius and in TBR. The main advantage of this algorithm was that no setting of DPM parameters was required in advance, since the proper setting of the only parameter that could significantly influence the segmentation results was automatically related to the uptake variance of the chosen ROI. Furthermore, the algorithm did not need any preliminary choice of the optimum number of classes to describe the ROIs within PET images and no assumption about the shape of the lesion and the uptake heterogeneity of the tracer was required.

The focus of this dissertation is to derive and demonstrate effective stochastic models for the speech recognition problem. Acoustic modeling for speech recognition typically involves representing the speech process within stochastic models. Modeling this high frequency time series effectively is a fundamental problem. This dissertation devised an objective function that relates the true speech distribution to its estimate. It is shown that through optimizing this function the speech process time series can be modeled without loss of information. The thesis proposes two such models that are developed to optimize the devised objective function. The first an acoustic model formulated for the speech with noise problem. The second a discriminately trained model consisting of optimal discriminant ML estimators. The first, a combination of recognizers that through a simple system fusion, combines multiple speech processes at the decision level. This is a stochastic modeling method devised to combine a parameterized spectral missing data, MD, theory based and a cepstral based speech process using a coupled hidden variable topology. In using a fused coupled hidden Markov model, HMM, topology, an optimal acoustic model is proposed that is inherently more robust than single process models under noisy conditions. The theoretical capability of this model is tested under both stationary and non stationary noise conditions. Under these test conditions the fused model has greater recognition accuracies than those of single process models. The second, formulated with a methodology that segments the acoustic space appropriately for discriminately trained models that optimize the devised objective function. This acoustic space is modeled with discriminant ML estimators formed with optimal decision boundaries using the large margin, support vector machine, SVM, learning method. These discriminately trained models maximize the entropy of the observation space and thereby are capable to model the speech process without loss. This is demonstrated experimentally with frame level classification error rates that are ∼ ≤ 3%.

Bayesian non-parametric (BNP) modeling has been developed and proven to be a powerful tool to analyze messy data with complex structures. Despite the increasing popularity of BNP modeling, it also faces challenges. One challenge is the estimation of the precision parameter in the Dirichlet process mixtures. In this study, we focus on a BNP growth curve model and investigate how non-informative prior, weakly informative prior, accurate informative prior, and inaccurate informative prior affect the model convergence, parameter estimation, and computation time. A simulation study has been conducted. We conclude that the non-informative prior for the precision parameter is less preferred because it yields a much lower convergence rate, and growth curve parameter estimates are not sensitive to informative priors.

This paper focuses on applications of Bayesian approaches to acoustic modeling for speech recognition and related speech-processing applications. Bayesian approaches have been widely studied in the fields of statistics and machine learning, and one of their advantages is that their generalization capability is better than that of conventional approaches (e.g., maximum likelihood). On the other hand, since inference in Bayesian approaches involves integrals and expectations that are mathematically intractable in most cases and require heavy numerical computations, it is generally difficult to apply them to practical speech recognition problems. However, there have been many such attempts, and this paper aims to summarize these attempts to encourage further progress on Bayesian approaches in the speech-processing field. This paper describes various applications of Bayesian approaches to speech processing in terms of the four typical ways of approximating Bayesian inferences, i.e., maximum a posteriori approximation, model complexity control using a Bayesian information criterion based on asymptotic approximation, variational approximation, and Markov chain Monte Carlo-based sampling techniques.

DNN-based acoustic modeling for speech recognition of native and foreign speakers

Acoustic modeling in large vocabulary continuous speech recognition systems is commonly done by building the models for subword units such as phonemes, syllables or senones. In recent years, various end-to-end systems using acoustic models built at grapheme or phoneme level have also been explored. These systems either require a lot of data and/or heavily rely on the use of language models or pronunciation dictionary for good recognition performance. With the intention of reducing the dependence on data or external models, we have explored the usage of phonetic features in building acoustic models for speech recognition. The phonetic features describe a sound based on the speech production mechanism in humans. Multi-label classification models are built for detection of phonetic features in a given speech signal. The detected phonetic features are used along with the acoustic features as input to models for phoneme identification. The effectiveness of the proposed approach is demonstrated on TIMIT and Wall Street Journal corpora. Performance improvement over other phoneme recognition studies using the phonetic features is obtained.