Phonological Constraints and Morphological Preprocessing for Grapheme-to-Phoneme Conversion

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 96–103, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Phonological Constraints and Morphological Preprocessing for Grapheme-to-Phoneme Conversion Vera Demberg School of Informatics University of Edinburgh Edinburgh, EH8 9LW, GB v.demberg@sms.ed.ac.uk Helmut Schmid IMS University of Stuttgart D-70174 Stuttgart schmid@ims.uni-stuttgart.de Gregor M¨ohler Speech Technologies IBM Deutschland Entwicklung D-71072 B¨oblingen moehler@de.ibm.com Abstract Grapheme-to-phoneme conversion (g2p) is a core component of any text-to-speech sys- tem. We show that adding simple syllab- ification and stress assignment constraints, namely ‘one nucleus per syllable’ and ‘one main stress per word’, to a joint n-gram model for g2p conversion leads to a dramatic improvement in conversion accuracy. Secondly, we assessed morphological pre- processing for g2p conversion. While mor- phological information has been incorpo- rated in some past systems, its contribution has never been quantitatively assessed for German. We compare the relevance of mor- phological preprocessing with respect to the morphological segmentation method, train- ing set size, the g2p conversion algorithm, and two languages, English and German. 1 Introduction Grapheme-to-Phoneme conversion (g2p) is the task of converting a word from its spelling (e.g. “Stern- anis¨ol”, Engl: star-anise oil) to its pronunciation (/"StERnPani:sPø:l/). Speech synthesis modules with a g2p component are used in text-to-speech (TTS) systems and can be be applied in spoken dialogue systems or speech-to-speech translation systems. 1.1 Syllabification and Stress in g2p conversion In order to correctly synthesize a word, it is not only necessary to convert the letters into phonemes, but also to syllabify the word and to assign word stress. The problems of word phonemization, syllabifica- tion and word stress assignment are inter-dependent. Information about the position of a syllable bound- ary helps grapheme-to-phoneme conversion. (Marc- hand and Damper, 2005) report a word error rate (WER) reduction of approx. 5 percentage points for English when the letter string is augmented with syl- labification information. The same holds vice-versa: we found that WER was reduced by 50% when run- ning our syllabifier on phonemes instead of letters (see Table 4). Finally, word stress is usually defined on syllables; in languages where word stress is as- sumed1 to partly depend on syllable weight (such as German or Dutch), it is important to know where ex- actly the syllable boundaries are in order to correctly calculate syllable weight. For German, (M¨uller, 2001) show that information about stress assignment and the position of a syllable within a word improve g2p conversion. 1.2 Morphological Preprocessing It has been argued that using morphological in- formation is important for languages where mor- phology has an important influence on pronuncia- tion, syllabification and word stress such as Ger- man, Dutch, Swedish or, to a smaller extent, also English (Sproat, 1996; M¨obius, 2001; Pounder and Kommenda, 1986; Black et al., 1998; Taylor, 2005). Unfortunately, these papers do not quantify the con- tribution of morphological preprocessing in the task. Important questions when considering the inte- gration of a morphological component into a speech 1This issue is controversial among linguists; …

Content truncated to 500 words