Latent Dirichlet allocation (LDA) is widely used for unsupervised topic modelling on sets of documents.No temporal information is used in the model.However, there is often a relationship between the corresponding topics of consecutive tokens.In this paper, we present an extension to LDA that uses a Markov chain to model temporal information.We use this new model for acoustic unit discovery from speech.As input tokens, the model takes a discretised encoding of speech from a vector quantised (VQ) neural network with 512 codes.The goal is then to map these 512 VQ codes to 50 phone-like units (topics) in order to more closely resemble true phones.In contrast to the base LDA, which only considers how VQ codes co-occur within utterances (documents), the Markov chain LDA additionally captures how consecutive codes follow one another.This extension leads to an increase in cluster quality and phone segmentation results compared to the base LDA.Compared to a recent vector quantised neural network approach that also learns 50 units, the extended LDA model performs better in phone segmentation but worse in mutual information.

Gaussian belief propagation (GaBP) is a message-passing algorithm that can be used to perform approximate inference on a pairwise Markov graph (MG) constructed from a multivariate Gaussian distribution in canonical parameterization. The output of GaBP is a set of approximate univariate marginals for each variable in the pairwise MG. An extension of GaBP (labeled GaBP-m), allowing for the approximation of higher-dimensional marginal distributions, was explored by Kamper et al. (2019). The idea is to create an MG in which each node is allowed to receive more than one variable. As in the univariate case, the multivariate extension does not necessarily converge in loopy graphs and, even if convergence occurs, is not guaranteed to provide exact inference. To address the problem of convergence, we consider a multivariate extension of the principle of node regularization proposed by Kamper et al. (2018). We label this algorithm slow GaBP-m (sGaBP-m), where the term 'slow' relates to the damping effect of the regularization on the message passing. We prove that, given sufficient regularization, this algorithm will converge and provide the exact marginal means at convergence, regardless of the way variables are assigned to nodes. The selection of the degree of regularization is addressed through the use of a heuristic, which is based on a tree representation of sGaBP-m. As a further contribution, we extend other GaBP variants in the literature to allow for higher-dimensional marginalization. We show that our algorithm compares favorably with these variants, both in terms of convergence speed and inference quality.

This paper is concerned with a multivariate extension of Gaussian message passing applied to pairwise Markov graphs (MGs). Gaussian message passing applied to pairwise MGs is often labeled Gaussian belief propagation (GaBP) and can be used to approximate the marginal of each variable contained in the pairwise MG. We propose a multivariate extension of GaBP (we label this GaBP-m) that can be used to estimate higher-dimensional marginals. Beyond the ability to estimate higher-dimensional marginals, GaBP-m exhibits better convergence behavior than GaBP, and can also provide more accurate univariate marginals. The theoretical results of this paper are based on an extension of the computation tree analysis conducted on univariate nodes to the multivariate case. The main contribution of this paper is the development of a convergence condition for GaBP-m that moves beyond the walk-summability of the precision matrix. Based on this convergence condition, we derived an upper bound for the number of iterations required for convergence of the GaBP-m algorithm. An upper bound on the dissimilarity between the approximate and exact marginal covariance matrices was established. We argue that GaBP-m is robust towards a certain change in variables, a property not shared by iterative solvers of linear systems, such as the conjugate gradient (CG) and preconditioned conjugate gradient (PCG) methods. The advantages of using GaBP-m over GaBP are also illustrated empirically.

The South African research community has strong individual interests in pattern recognition and machine learning, but to date has had limited interactions with the worldwide multimedia research community. In an attempt to redress this, this workshop aims to introduce a selection of South African researchers to the multimedia community, and expose the multimedia community to a range of multimedia-related work, primarily from South Africa.

Download bandwidth is often a limiting factor in the remote sensing image acquisition chain. An onboard image quality assessment system could optimise use of available downlink time by prioritising high quality images for download. An image quality assessment model was developed based on 3 degradation features: cloud cover, additive noise and the defocus extent of the telescope. For training data more than 18000 subjective evaluations of 484 unique degraded images were collected over the internet. A neural network model and compound spline model were trained and compared. The spline model achieved a linear correlation coefficient of 96,5%.

Static handwritten scripts are available only as images on documents and by definition do not contain dynamic information. This study is about extracting dynamic information from a static handwritten script, specifically the sequence of pen positions that created the script. We assume that a dynamic representative of the static image is available (a different version typically obtained during an earlier registration process). A hidden Markov model (HMM) of the static image is compared with the dynamic representative to extract the dynamic information from the static image.

Static signatures originate as handwritten images on documents and by definition do not contain any dynamic information. This lack of information makes static signature verification systems significantly less reliable than their dynamic counterparts. This study involves extracting dynamic information from static images, specifically the pen trajectory while the signature was created. We assume that a dynamic version of the static image is available (typically obtained during an earlier registration process). We then derive a hidden Markov model from the static image and match it to the dynamic version of the image. This match results in the estimated pen trajectory of the static image.

This paper presents a purely statistical method for the automatic syllabification of speech. A hierarchical HMM structure is used to implement a purely acoustical model based on the phonotactic constraints found in the English language. A well-defined DTW distance measure is presented for measuring and reporting syllabification results. We achieve a token error rate of 20.3 % with a 42 ms average boundary error on a relatively large set of data. This compares well with previous knowledge- and statistically based methods.

The field of topic spotting in conversational speech deals with the problem of identifying “interesting” conversations or speech extracts amongst large volumes of speech data. In this research, two phoneme-based topic spotting systems were evaluated on the Switchboard Corpus. Experiments [1, 2] on the OGI Corpus suggested that the new Stochastic Method for the Automatic Recognition of Topics (SMART) yields a large improvement over the existing Euclidean Nearest Wrong Neighbours (ENWN) algorithm, which had outperformed competing systems in evaluations [3, 4]. However, the small amount of data available for these experiments meant that more rigorous testing was required. We reimplemented the algorithm to run on the larger Switchboard Corpus, and report an improvement of SMART over ENWN characterised by a 35.8% reduction in ROC (receiver operating characteristic) error area. Statistical significance was demonstrated using a modified version of the McNemar test.

Higher-order Markov models are more powerful than first-order models, but suffer from an exponential increase in model parameters with order, which leads to data scarcity problems during training. A more efficient approach is to use mixed-order Markov models, which model data sequences with contexts of different lengths. This study proposes two algorithms for inferring mixed-order Markov chains and hidden Markov models (HMMs), respectively. The basis of these algorithms is the prediction suffix tree (PST), an efficient representation of a mixed-order Markov chain. The smallest encoded context tree (SECT) algorithm constructs PSTs from data, based on the minimum description length principle. It has no user-specifiable parameters to tune, and will expand the depth of the resulting PST as far as the data set allows it, making it a self-bounded algorithm. It is also faster than the original PST inference algorithm. The hidden SECT algorithm replaces the underlying Markov chain of an HMM with a prediction suffix tree, which is inferred using SECT. The algorithm is efficient and integrates well with standard techniques. The properties of the SECT and hidden SECT algorithms are verified on synthetic data. The hidden SECT algorithm is also compared with a fixed-order HMM training algorithm on an automatic language recognition task, where the resulting mixed-order HMMs are shown to be smaller and train faster than the fixed-order models, for similar classification accuracies.

I, the undersigned, hereby declare that the work contained in this dissertation is my own original work and that I have not previously in its entirety or in part submitted it at any university for a degree. Signature: Date: ii Currently, first-order hidden Markov models (HMMs) form the backbone around which most automatic speech processing applications are built. Their higher-order extensions are known to be more powerful, but, due to their complexity and computational demands, they are seldomly used. It is the purpose of this work to advance their application In this work we unify HMMs of all orders by deriving and proving the ORder rEDucing (ORED) algorithm. This algorithm will reduce any higher-order HMM (also mixed-order) to an equivalent first-order representation. This makes it possible to process any higher-order HMM using known first-order algorithms, thereby

A speaker independent technique for identifying stops in continuous speech is described. The stops are modeled with continuous hidden Markov models (CHMMs) as consisting of several well-defined segments: silence, voicing, a release and aspiration. These models are capable of performing two tasks. The first is the classification of an unknown stop and the second is to obtain a fine transcription of the stop into its segments. Features pertinent to stop recognition are obtained from the segment boundaries and are used together with the model scores in a nonparametric probability density function (PDF) estimator to identify unknown stop consonants. A recognition rate of 84% was achieved on stops occurring in vowel-stop-vowel clusters that were taken from continuous speech.>