The collection of large speech databases is not a trivial task (if done properly). It is not always possible to collect, segment and annotate large databases for every task or language. It is also often the case that there are imbalances in the databases, as a result of little data being available for a specific subset of individuals. An example of one such imbalance is the fact that there are often more male speakers than female speakers (or vice-versa). If there are, for example, far fewer female speakers than male speakers, then the recognizers will tend to work poorly for female speakers (as compared to performance for male speakers). This thesis focuses on using Bayesian and discriminative training algorithms to improve continuous speech recognition systems in scenarios where there is a limited amount of training data available. The research reported in this thesis can be divided into three categories: • Overspecialization is characterized by good recognition performance for the data used during training, but poor recognition performance for independent testing data. This is a problem when too little data is available for training purposes. Methods of reducing overspecialization in the minimum classification error algo¬rithm are therefore investigated. • Development of new Bayesian and discriminative adaptation/training techniques that can be used in situations where there is a small amount of data available. One example here is the situation where an imbalance in terms of numbers of male and female speakers exists and these techniques can be used to improve recognition performance for female speakers, while not decreasing recognition performance for the male speakers. • Bayesian learning, where Bayesian training is used to improve recognition perfor¬mance in situations where one can only use the limited training data available. These methods are extremely computationally expensive, but are justified by the improved recognition rates for certain tasks. This is, to the author's knowledge, the first time that Bayesian learning using Markov chain Monte Carlo methods have been used in hidden Markov model speech recognition. The algorithms proposed and reviewed are tested using three different datasets (TIMIT, TIDIGITS and SUNSpeech), with the tasks being connected digit recognition and con¬tinuous speech recognition. Results indicate that the proposed algorithms improve recognition performance significantly for situations where little training data is avail¬able. / Thesis (PhD (Electronic Engineering))--University of Pretoria, 2006. / Electrical, Electronic and Computer Engineering / unrestricted

Speech recognition systems have been developed for the major languages of the world, yet for the majority of languages there are currently no large vocabulary continuous speech recognition (LVCSR) systems. The development of an LVCSR system for a new language is very costly, mainly because a large speech database has to be compiled to robustly capture the acoustic characteristics of the new language. This thesis investigates techniques that enable the re-use of acoustic information from a source language, in which a large amount of data is available, in implementing a system for a new target language. The assumption is that too little data is available in the target language to train a robust speech recognition system on that data alone, and that use of acoustic information from a source language can improve the performance of a target language recognition system. Strategies for cross-language use of acoustic information are proposed, including training on pooled source and target language data, adaptation of source language models using target language data, adapting multilingual models using target language data and transforming source language data to augment target language data for model training. These strategies are allied with Bayesian and transformation-based techniques, usually used for speaker adaptation, as well as with discriminative learning techniques, to present a framework for cross-language re-use of acoustic information. Extensions to current adaptation techniques are proposed to improve the performance of these techniques specifically for cross-language adaptation. A new technique for transformation-based adaptation of variance parameters and a cost-based extension of the minimum classification error (MCE) approach are proposed. Experiments are performed for a large number of approaches from the proposed framework for cross-language re-use of acoustic information. Relatively large amounts of English speech data are used in conjunction with smaller amounts of Afrikaans speech data to improve the performance of an Afrikaans speech recogniser. Results indicate that a significant reduction in word error rate (between 26% and 50%, depending on the amount of Afrikaans data available) is possible when English acoustic data is used in addition to Afrikaans speech data from the same database (i.e both sets of data were recorded under the same c`12onditions and the same labelling process was used). For same-database experiments, best results are achieved for approaches that train models on pooled source and target language data and then perform further adaptation of the models using Bayesian or discriminative techniques on target language data only. Experiments are also performed to evaluate the use of English data from a different database than the Afrikaans data. Peak reductions in word error rate of between 16% and 35% are delivered, depending on the amount of Afrikaans data available. Best results are achieved for an approach that performs a simple transformation of source model parameters using target language data, and then performs Bayesian adaptation of the transformed model on target language data. / Thesis (PhD (Electrical, Electronic and Computer Engineering))--University of Pretoria, 2006. / Electrical, Electronic and Computer Engineering / unrestricted