Automatic Detection of Syllable Boundaries Combining the Advantages of Treebank and Bracketed Corpora Training

Automatic Detection of Syllable Boundaries Combining the Advantages of Treebank and Bracketed Corpora Training Karin Müller Institut für Maschinelle Sprachverarbeitung University of Stuttgart Azenbergstrasse 12 D-70174 Stuttgart, Germany karin.mueller@ims.uni-stuttgart.de Abstract An approach to automatic detection of syllable boundaries is presented. We demonstrate the use of several manu- ally constructed grammars trained with a novel algorithm combining the advan- tages of treebank and bracketed corpora training. We investigate the effect of the training corpus size on the perfor- mance of our system. The evaluation shows that a hand-written grammar per- forms better on finding syllable bound- aries than does a treebank grammar. 1 Introduction In this paper we present an approach to super- vised learning and automatic detection of sylla- ble boundaries. The primary goal of the paper is to demonstrate that under certain conditions treebank and bracketed corpora training can be combined by exploiting the advantages of the two methods. Treebank training provides a method of unambiguous analyses whereas bracketed corpora training has the advantage that linguistic knowl- edge can be used to write linguistically motivated grammars. In text-to-speech (TTS) systems, like those de- scribed in Sproat (1998), the correct pronuncia- tion of unknown or novel words is one of the biggest problems. In many TTS systems large pronunciation dictionaries are used. However, the lexicons are finite and every natural language has productive word formation processes. The German language for example is known for its extensive use of compounds. A TTS system needs a module where the words converted from graphemes to phonemes are syllabified before they can be further processed to speech. The placement of the correct syllable boundary is es- sential for the application of phonological rules (Kahn, 1976; Blevins, 1995). Our approach of- fers a machine learning algorithm for predicting syllable boundaries. Our method builds on two resources. The first resource is a series of context-free gram- mars (CFG) which are either constructed manu- ally or extracted automatically (in the case of the treebank grammar) to predict syllable boundaries. The different grammars are described in section 4. The second resource is a novel algorithm that aims to combine the advantages of treebank and bracketed corpora training. The obtained proba- bilistic context-free grammars are evaluated on a test corpus. We also investigate the influence of the size of the training corpus on the performance of our system. The evaluation shows that adding linguistic in- formation to the grammars increases the accuracy of our models. For instance, we coded the knowl- edge that (i) consonants in the onset and coda are restricted in their distribution, and (ii) the position inside of the word plays an important role. Fur- thermore, linguistically motivated grammars only need a small size of training corpus to achieve high accuracy and even out-perform the treebank grammar trained on the largest training corpus. The remainder of the paper is organized as fol- lows. Section 2 refers to treebank training. In section 3 we introduce the combination of tree- [ f On Onset O Nucleus R Cod …

Content truncated to 500 words