2007: Vol 98 No 4

Download the this issue (complete journal, excluding cover) (~2.4 MB).

Jump to:

1. A Monophone Speech Generation System
2. Effort and Accuracy during Language Resource Generation: A Pronunciation Prediction Case Study
3. Speech Rate Normalization used to Improve Speaker Verification
4. Statistical Translation with Scarce Resources: A South African Case Study
5. Text-Based Language Identification for South African Languages

A Monophone Speech Generation System by G. klompje and T.R. Niesler
Abstract: Current speech synthesis systems generally require large and carefully annotated speech corpora for their development. However, for many languages these resources are not available. This paper describes a speech generation algorithm based on monophone subword units for minimal reliance on such databases. The system is based on the source-filter speech production framework, and includes a linear prediction based vocal tract model as well as an excitation model. An interpolation algorithm is presented to allow coarticulation between monophone units to be modelled. The excitation model includes a method for dealing with voiced and partiallyvoiced sounds based on a Gaussianity measure applied to the excitation spectrum. Promising first results were obtained when evaluating the intelligibility of the developed system’s South African English speech output using the modified rhyme test and semantically unpredictable sentences.
Download Paper (~1 Mb)

Effort and Accuracy during Language Resource Generation: A Pronunciation Prediction Case Study by M. Davel and E. Barnard
Abstract: When developing a language resource, there is generally a trade-off between the amount of effort invested in the resource creation process and the quality of the resulting resource. We argue that, in the developing world with its many resource-scarce languages, a ‘usable’ resource in multiple languages may be more valuable than a highly accurate resource for one language only. From this perspective we investigate the resource validation process – determining whether a resource is sufficiently accurate – using the creation of a pronunciation dictionary as case study. We show that the amount of effort required to validate a 20,000-word pronunciation dictionary can be reduced substantially by employing appropriate computational tools, when compared to both a fully manual validation process and a competing automatic process.
Download Paper (~261 kb)

Speech Rate Normalization used to Improve Speaker Verification by C.J. van Heerden and E. Barnard
Abstract: A novel approach to speech rate normalization is presented. Models are constructed to model the way in which speech rate variation of a specific speaker influences the duration of phonemes. The models are evaluated in two ways. Firstly, the mean square error in phoneme duration based on our normalization is compared to the same error when such normalization is not applied. The second evaluation uses the durations of context-dependent phonemes in speaker verification. Both methods show that this approach to normalization is indeed effective to counteract the effect of variable speaking rates.
Download Paper (~0.5 Mb)

Statistical Translation with Scarce Resources: A South African Case Study by R.S.M. Kato and E. Barnard
Abstract: Statistical machine translation techniques offer great promise for the development of automatic translation systems. However, the realization of this potential requires the availability of significant amounts of parallel bilingual texts. This paper reports on an attempt to reduce the amount of text that is required to obtain an acceptable translation system, through the use of active and semisupervised learning. Systems were built using resources collected from South African government websites and the results evaluated using a standard automatic evaluation metric (BLEU). We show that significant improvements in translation quality can be achieved with very limited parallel corpora, and that both active learning and semi-supervised learning are useful in this context.
Download Paper (~250 kb)

Text-Based Language Identification for South African Languages by Gerrit Botha, Victor Zimu and Etienne Barnard
Abstract: We investigate the performance of text-based language identification systems on the 11 official languages of South Africa, when n-gram statistics are used as features for classification. In particular, we compare support vector machines, likelihood and frequency difference-based classifiers on different amounts of input text and for various values of n. With as few as 15 words of input text, reliable language identification is possible. Although the support vector macine is generally more accurate as classifier, the additional computational complexity of training this classifier may not be justified in light of the importance of using a large value for n.
Download Paper (469 kb)