si it ty y o of f P Pr re et to or ri ia aTo Carina, who makes every moment in life special.

Discriminative training of hidden Markov models (HMMs) using minimum classification error training (MCE) has been shown to work well for certain speech recognition applications. MCE is, however, somewhat prone to overspecialization. This study investigates various techniques which improve performance and generalization of the MCE algorithm. Improvements of up to 10% in relative error rate on the test set are achieved for the TIMIT dataset.

The African Speech Technology project is developing telephone speech databases for five of South Africa’s eleven official languages, i.e. South African English, Afrikaans, and three African languages, Zulu, Xhosa, and Southern Sotho. These databases will be fully transcribed – orthographically and phonetically – and will be used for the training and testing of phoneme-based, speaker-independent speech recognition systems. This paper describes the design and contents of the speech corpus that is currently being collected over both mobile and fixed networks. In particular language coverage is discussed within the framework of the multilingual character of the South African population. Some language-specific differences with regards to the contents of the different databases are noted. Methods and tools applied in the acquisition of phonetic information are discussed.

This paper presents a Bayesian approach to learning for HMMs in speech recognition. The implementation of Bayesian learning for HMMs in speech recognition is discussed, including the requirement of maintaining the original HMM constraints, choice of prior and utterance recognition. This work shows that the Bayesian learning approach can be successfully applied to complex models when the amount of training data is small.

Techniques are investigated that use acoustic information from existing source language databases to implement automatic speech recognition (ASR) systems for new target languages. The assumption is that the amount of target language data available is too little for the training of a robust ASR system. Strategies for cross-language use of acoustic information are evaluated which include (i) training on pooled source and target language data, (ii) adapting source language models using target language data, (iii) adapting models trained on pooled source and target language using target language data only and (iv) transforming source language data to augment target language data for model training. These strategies are allied with Bayesian and transformation-based techniques to present a framework for cross-language reuse of acoustic information. Experiments are performed for a large number of approaches from the framework, using relatively large amounts of English speech data from either a separate database or from the same-database as 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 is achievable (between 14% and 48% for experiments), depending on the amount of target language data available.

A surveillance system which automatically classifies aircraft can be very useful, especially with the problem of congestion at large airports. This paper evaluates the performance of correlation- and feature-based classifiers on a set of four simulated radar targets over a wide range of target orientations. Experiments are performed for a range of radar bandwidths in order to determine the effect of radar bandwidth on the relative classification performance. Only 1D radar range profiles are considered since it is assumed that the aircraft are classified using few profiles and the orientation of the aircraft in each profile is known only approximately. The results suggest that feature-based classifiers outperform correlation-based classifiers and that classification performance is highly dependent on the orientation of the aircraft, but that accurate classification of approaching aircraft is possible.

In this paper we present a parallel implementation of a well-known heuristic optimisation algorithm (the downhill simplex algorithm developed by Nelder and Mead in 1965) which is well suited for unconstrained optimisation. We present the sequential algorithm as well as the parallel algorithm which we used to generate numerical results. They include numerical results of experiments on neural networks and a test suite of functions which demonstrate the parallel algorithm's increased robustness and convergence rate for high-dimensional problems compared to the sequential algorithm. © 1998 John Wiley & Sons, Ltd.

In modern day pattern recognition, neural nets are used extensively. General use of a feedforward neural net consists of a training phase followed by a classification phase. Classification of an unknown test vector is very fast and only consists of the propagation of the test vector through the neural net. Training involves an optimization procedure and is very time-consuming since a feasible local minimum is sought in high-dimensional weight space. In this paper we present an analysis of a parallel implementation of the backpropagation training algorithm using conjugate gradient optimization for a three-layered, feedforward neural network, on the KSR1 parallel shared-memory machine. We implement two parallel neural net training versions on the KSR1, one using native code, the other using P4, a library of macros and functions. A speedup model is presented which we use to clarify our experimental results. We identify the general requirements which render the parallel implementation useful, compared to the sequential execution of the same neural net training procedure. We determine the usefulness of a library of functions (such as P4) developed to ease the task of the programmer. Using experimental results we further identify the limits in processor utilization for our parallel training algorithm.

The ability of neural net classifiers to deal with a priori information is investigated. For this purpose, backpropagation classifiers are trained with data from known distributions with variable a priori probabilities, and their performance on separate test sets is evaluated. It is found that backpropagation employs a priori information in a slightly suboptimal fashion, but this does not have serious consequences on the performance of the classifier. Furthermore, it is found that the inferior generalization that results when an excessive number of network parameters are used can (partially) be ascribed to this suboptimality.