Supervised learning using Hidden Markov Models has been used to train acoustic models for automatic speech recognition for several years. Typically clean transcriptions form the basis for this training regimen. However, results have shown that using sources of readily available transcriptions, which can be erroneous at times (e.g., closed captions) do not degrade the performance significantly. This work analyzes the effects of mislabeled data on recognition accuracy. For this purpose, the training is performed using manually corrupted training data and the results are observed on three different databases: TIDigits, Alphadigits and SwitchBoard. For Alphadigits, with 16% of data mislabeled, the performance of the system degrades by 12% relative to the baseline results. For a complex task like SWITCHBOARD, at 16% mislabeled training data, the performance of the system degrades by 8.5% relative to the baseline results. The training process is more robust to mislabeled data because the Gaussian mixtures that are used to model the underlying distribution tend to cluster around the majority of the correct data. The outliers (incorrect data) do not contribute significantly to the reestimation process.

In Natural Language Processing (NLP), speech and text are parsed and generated with language models and parser models, and translated with translation models. Each model contains a set of numerical parameters which are found by applying a suitable training algorithm to a set of training data. Many such training algorithms are instances of the Expectation-Maximization (EM) algorithm. In [BSV15], a generic EM algorithm for NLP is described. This work presents a particular speech model, the Hidden Markov model, and its standard training algorithm, the Baum-Welch algorithm. It is then shown that the Baum-Welch algorithm is an instance of the generic EM algorithm introduced by [BSV15], from which follows that all statements about the generic EM algorithm also apply to the Baum-Welch algorithm, especially its correctness and convergence properties.

DNA can be represented abstrzctly as a language with only four nucleotides represented by the letters A, C, G, and T, yet the arrangement of those four letters plays a major role in determining the development of an organism. Understanding the signi cance of certain arrangements of nucleotides can unlock the secrets of how the genome achieves its essential functionality. Regions of DNA particularly enriched with cytosine (C nucleotides) and guanine (G nucleotides), especially the CpG di-nucleotide, are frequently associated with biological function related to gene expression, and concentrations of CpGs referred to as \CpG islands" are known to collocate with regions upstream from gene coding sequences within the promoter region. The pattern of occurrence of these nucleotides, relative to adenine (A nucleotides) and thymine (T nucleotides), lends itself to analysis by machine-learning techniques such as Hidden Markov Models (HMMs) to predict the areas of greater enrichment. HMMs have been applied to CpG island prediction before, but often without an awareness of how the outcomes are a ected by the manner in which the HMM is applied. Two main ndings of this study are: 1. The outcome of a HMM is highly sensitive to the setting of the initial probability estimates. 2. Without the appropriate software techniques, HMMs cannot be applied e ectively to large data such as whole eukaryotic chromosomes. Both of these factors are rarely considered by users of HMMs, but are critical to a successful application of HMMs to large DNA sequences. In fact, these shortcomings were discovered through a close examination of published results of CpG island prediction using HMMs, and without being addressed, can lead to an incorrect implementation and application of HMM theory. A rst-order HMM is developed and its performance compared to two other historical methods, the Takai and Jones method and the UCSC method from the University of California Santa Cruz. The HMM is then extended to a second-order to acknowledge that pairs of nucleotides de ne CpG islands rather than single nucleotides alone, and the second-order HMM is evaluated in comparison to the other methods. The UCSC method is found to be based on properties that are not related to CpG islands, and thus is not a fair comparison to the other methods. Of the other methods, the rst-order HMM method and the Takai and Jones method are comparable in the tests conducted, but the second-order HMM method demonstrates superior predictive capabilities. However, these results are valid only when taking into consideration the highly sensitive outcomes based on initial estimates, and nding a suitable set of estimates that provide the most appropriate results. The rst-order HMM is applied to the problem of producing synthetic data that simulates the characteristics of a DNA sequence, including the speci ed presence of CpG islands, based on the model parameters of a trained HMM. HMM analysis is applied to the synthetic data to explore its delity in generating data with similar characteristics, as well as to validate the predictive ability of an HMM. Although this test fails to i meet expectations, a second test using a second-order HMM to produce simulated DNA data using frequency distributions of CpG island pro les exhibits highly accurate predictions of the pre-speci ed CpG islands, con- rming that when the synthetic data are appropriately structured, an HMM can be an accurate predictive tool. One outcome of this thesis is a set of software components (CpGID 2.0 and TrackMap) capable of ef- cient and accurate application of an HMM to genomic sequences, together with visualization that allows quantitative CpG island results to be viewed in conjunction with other genomic data. CpGID 2.0 is an adaptation of a previously published software component that has been extensively revised, and TrackMap is a companion product that works with the results produced by the CpGID 2.0 program. Executing these components allows one to monitor output aspects of the computational model such as number and size of the predicted CpG islands, including their CG content percentage and level of CpG frequency. These outcomes can then be related to the input values used to parameterize the HMM.

A speaker independent continuous speech phoneme recognition and segmentation system is presented. We discuss the training and recognition phases of the phoneme recognition system as well as a detailed description of the integrated elements. The Hidden Markov Model (HMM) based phoneme models are trained using the Baum-Welch re-estimation procedure. Recognition and segmentation of the phonemes in the continuous speech is performed by a Segmental Viterbi Search on a Segmental Ergodic HMM for the phoneme states. We describe in detail the three phases of the phoneme joint recognition and segmentation system. First, the extraction of the Mel-Frequency Cepstral Coefficients (MFCC) and the corresponding Delta and Delta Log Power coefficients is described. Second, we describe the operation of the Baum-Welch re-estimation procedure for the training of the phoneme HMM models, including the K-Means and the Expectation-Maximization (EM) clustering algorithms used for the initialization of the Baum-Welch algorithm. Additionally, we describe the structural framework of - and the recognition procedure for - the ergodic Segmental HMM for the phoneme segmentation and recognition. We include test and simulation results for each of the individual systems integrated into the phoneme recognition system and finally for the phoneme recognition/segmentation system as a whole. / Master of Science

Since the 1950s, hidden Markov models (HMMS) have seen widespread use in electrical engineering. Foremost has been their use in speech processing, pattern recognition, artificial intelligence, queuing theory, and communications theory. However, recent years have witnessed a renaissance in the application of HMMs to the analysis and simulation of digital communication systems. Typical applications have included signal estimation, frequency tracking, equalization, burst error characterization, and transmit power control. Of special significance to this thesis, however, has been the use of HMMs to model fading channels typical of wireless communications. This variegated use of HMMs is fueled by their ability to model time-varying systems with memory, their ability to yield closed form solutions to otherwise intractable analytic problems, and their ability to help facilitate simple hardware and/or software based implementations of simulation test-beds. The aim of this thesis is to employ and exploit hidden Markov fading models within an iterative (turbo) decoding framework. Of particular importance is the problem of channel estimation, which is vital for realizing the large coding gains inherent in turbo coded schemes. This thesis shows that a Markov fading channel (MFC) can be conceptualized as a trellis, and that the transmission of a sequence over a MFC can be viewed as a trellis encoding process much like convolutional encoding. The thesis demonstrates that either maximum likelihood sequence estimation (MLSE) algorithms or maximum <I> a posteriori</I> (MAP) algorithms operating over the trellis defined by the MFC can be used for channel estimation. Furthermore, the thesis illustrates sequential and decision-directed techniques for using the aforementioned trellis based channel estimators <I>en masse</I> with an iterative decoder. / Master of Science

In Natural Language Processing (NLP), speech and text are parsed and generated with language models and parser models, and translated with translation models. Each model contains a set of numerical parameters which are found by applying a suitable training algorithm to a set of training data. Many such training algorithms are instances of the Expectation-Maximization (EM) algorithm. In [BSV15], a generic EM algorithm for NLP is described. This work presents a particular speech model, the Hidden Markov model, and its standard training algorithm, the Baum-Welch algorithm. It is then shown that the Baum-Welch algorithm is an instance of the generic EM algorithm introduced by [BSV15], from which follows that all statements about the generic EM algorithm also apply to the Baum-Welch algorithm, especially its correctness and convergence properties.:1 Introduction 1.1 N-gram models 1.2 Hidden Markov model 2 Expectation-maximization algorithms 2.1 Preliminaries 2.2 Algorithmic skeleton 2.3 Corpus-based step mapping 2.4 Simple counting step mapping 2.5 Regular tree grammars 2.6 Inside-outside step mapping 2.7 Review 3 The Hidden Markov model 3.1 Forward and backward algorithms 3.2 The Baum-Welch algorithm 3.3 Deriving the Baum-Welch algorithm 3.3.1 Model parameter and countable events 3.3.2 Tree-shaped hidden information 3.3.3 Complete-data corpus 3.3.4 Inside weights 3.3.5 Outside weights 3.3.6 Complete-data corpus (cont.) 3.3.7 Step mapping 3.4 Review Appendix A Elided proofs from Chapter 3 A.1 Proof of Lemma 3.8 A.2 Proof of Lemma 3.9 B Formulary for Chapter 3 Bibliography

This thesis explored the possibilities of using a Hidden Markov Model approach for multi-target localisation in an urban environment, with observations generated from Wi-Fi sensors. The area is modelled as a network of nodes and arcs, where the arcs represent sidewalks in the area and constitutes the hidden states in the model. The output of the model is the expected amount of people at each road segment throughout the day. In addition to this, two methods for analyzing the impact of events in the area are proposed. The first method is based on a time series analysis, and the second one is based on the updated transition matrix using the Baum-Welch algorithm. Both methods reveal which road segments are most heavily affected by a surge of traffic in the area, as well as potential bottleneck areas where congestion is likely to have occurred. / I det här examensarbetet har lokalisering av gångtrafikanter med hjälp av Hidden Markov Models utförts. Lokaliseringen är byggd på data från Wi-Fi sensorer i ett område i Stockholm. Området är modellerat som ett graf-baserat nätverk där linjerna mellan noderna representerar möjliga vägar för en person att befinna sig på. Resultatet för varje individ är aggregerat för att visa förväntat antal personer på varje segment över en hel dag. Två metoder för att analysera hur event påverkar området introduceras och beskrivs. Den första är baserad på tidsserieanalys och den andra är en maskinlärningsmetod som bygger på Baum-Welch algoritmen. Båda metoderna visar vilka segment som drabbas mest av en snabb ökning av trafik i området och var trängsel är troligt att förekomma.

This thesis provides an in-depth analyze of expressed sequence tags (EST) that represent pieces of eukaryotic mRNA by using unsupervised hidden Markov model (HMM). ESTs are short nucleotide sequences that are used primarily for rapid identificationof new genes with potential coding regions (CDS). ESTs are made by sequencing on double-stranded cDNA and the synthesizedESTs are stored in digital form, usually in FASTA format. Since sequencing is often randomized and that parts of mRNA contain non-coding regions, some ESTs will not represent CDS.It is desired to remove these unwanted ESTs if the purpose is to identifygenes associated with CDS. Application of stochastic HMM allow identification of region contents in a EST. Softwares like ESTScanuse HMM in which a training of the HMM is done by supervised learning with annotated data. However, because there are not always annotated data at hand this thesis focus on the ability to train an HMM with unsupervised learning on data containing ESTs, both with and without CDS. But the data used for training is not annotated, i.e. the regions that an EST consists of are unknown. In this thesis a new HMM is introduced where the parameters of the HMM are in focus so that they are reasonablyconsistent with biologically important regionsof an mRNA such as the Kozak sequence, poly(A)-signals and poly(A)-tails to guide the training and decoding correctly with ESTs to proper statesin the HMM. Transition probabilities in the HMMhas been adapted so that it represents the mean length and distribution of the different regions in mRNA. Testing of the HMM's specificity and sensitivityhave been performed via BLAST by blasting each EST and compare the BLAST results with the HMM prediction results.A regression analysis test shows that the length of ESTs used when training the HMM is significantly important, the longer the better. The final resultsshows that it is possible to train an HMM with unsupervised machine learning but to be comparable to supervised machine learning as ESTScan, further expansion of the HMM is necessary such as frame-shift correction of ESTs byimproving the HMM's ability to choose correctly positioned start codons or nucleotides. Usually the false positive results are because of incorrectly positioned start codons leadingto too short CDS lengths. Since no frame-shift correction is implemented, short predicted CDS lengths are not acceptable and is hence not counted as coding regionsduring prediction. However, when there is a lack of supervised models then unsupervised HMM is a potential replacement with stable performance and able to be adapted forany eukaryotic organism.

This dissertation explores a unique automated approach to map one phoneme set to another, based on the acoustic distances between the individual phonemes. Although the focus of this investigation is on cross-language applications, this automated approach can be extended to same-language but different-database applications as well. The main goal of this investigation is to be able to use the data of a source language, to train the initial acoustic models of a target language for which very little speech data may be available. To do this, an automatic technique for mapping the phonemes of the two data sets must be found. Using this technique, it would be possible to accelerate the development of a speech recognition system for a new language. The current research in the cross-language speech recognition field has focused on manual methods to map phonemes. This investigation has considered an English-to-Afrikaans phoneme mapping, as well as an Afrikaans-to-English phoneme mapping. This has been previously applied to these language instances, but utilising manual phoneme mapping methods. To determine the best phoneme mapping, different acoustic distance measures are compared. The distance measures that are considered are the Kullback-Leibler measure, the Bhattacharyya distance metric, the Mahalanobis measure, the Euclidean measure, the L2 metric and the Jeffreys-Matusita distance. The distance measures are tested by comparing the cross-database recognition results obtained on phoneme models created from the TIMIT speech corpus and a locally-compiled South African SUN Speech database. By selecting the most appropriate distance measure, an automated procedure to map phonemes from the source language to the target language can be done. The best distance measure for the mapping gives recognition rates comparable to a manual mapping process undertaken by a phonetic expert. This study also investigates the effect of the number of Gaussian mixture components on the mapping and on the speech recognition system’s performance. The results indicate that the recogniser’s performance increases up to a limit as the number of mixtures increase. In addition, this study has explored the effect of excluding the Mel Frequency delta and acceleration cepstral coefficients. It is found that the inclusion of these temporal features help improve the mapping and the recognition system’s phoneme recognition rate. Experiments are also carried out to determine the impact of the number of HMM recogniser states. It is found that single-state HMMs deliver the optimum cross-language phoneme recognition results. After having done the mapping, speaker adaptation strategies are applied on the recognisers to improve their target-language performance. The models of a fully trained speech recogniser in a source language are adapted to target-language models using Maximum Likelihood Linear Regression (MLLR) followed by Maximum A Posteriori (MAP) techniques. Embedded Baum-Welch re-estimation is used to further adapt the models to the target language. These techniques result in a considerable improvement in the phoneme recognition rate. Although a combination of MLLR and MAP techniques have been used previously in speech adaptation studies, the combination of MLLR, MAP and EBWR in cross-language speech recognition is a unique contribution of this study. Finally, a data pooling technique is applied to build a new recogniser using the automatically mapped phonemes from the target language as well as the source language phonemes. This new recogniser demonstrates moderate bilingual phoneme recognition capabilities. The bilingual recogniser is then further adapted to the target language using MAP and embedded Baum-Welch re-estimation techniques. This combination of adaptation techniques together with the data pooling strategy is uniquely applied in the field of cross-language recognition. The results obtained using this technique outperform all other techniques tested in terms of phoneme recognition rates, although it requires a considerably more time consuming training process. It displays only slightly poorer phoneme recognition than the recognisers trained and tested on the same language database. / Dissertation (MEng (Computer Engineering))--University of Pretoria, 2006. / Electrical, Electronic and Computer Engineering / unrestricted

Probabilistic graphical modeling is a framework which can be used to succinctly<br>represent multivariate probability distributions of time series in terms of each time<br>series’s dependence on others. In general, it is computationally prohibitive to sta-<br>tistically infer an arbitrary model from data. However, if we constrain the model to<br>have a tree topology, the corresponding learning algorithms become tractable. The<br>expressive power of tree-structured distributions are low, since only n − 1 dependen-<br>cies are explicitly encoded for an n node tree. One way to improve the expressive<br>power of tree models is to combine many of them in a mixture model. This work<br>presents and uses simulations to validate extensions of the standard mixtures of trees<br>model for i.i.d data to the setting of time series data. We also consider the setting<br>where the tree mixture itself forms a hidden Markov chain, which could be better<br>suited for approximating time-varying seasonal data in the real world. Both of these<br>are evaluated on artificial data sets.<br><br>