On the Syllabification of Phonemes

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 308–316, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics On the Syllabification of Phonemes Susan Bartlett† and Grzegorz Kondrak† and Colin Cherry‡ †Department of Computing Science ‡Microsoft Research University of Alberta One Microsoft Way Edmonton, AB, T6G 2E8, Canada Redmond, WA, 98052 {susan,kondrak}@cs.ualberta.ca colinc@microsoft.com Abstract Syllables play an important role in speech synthesis and recognition. We present sev- eral different approaches to the syllabifica- tion of phonemes. We investigate approaches based on linguistic theories of syllabification, as well as a discriminative learning technique that combines Support Vector Machine and Hidden Markov Model technologies. Our experiments on English, Dutch and German demonstrate that our transparent implemen- tation of the sonority sequencing principle is more accurate than previous implemen- tations, and that our language-independent SVM-based approach advances the current state-of-the-art, achieving word accuracy of over 98% in English and 99% in German and Dutch. 1 Introduction Syllabification is the process of dividing a word into its constituent syllables. Although some work has been done on syllabifying orthographic forms (M¨uller et al., 2000; Bouma, 2002; Marchand and Damper, 2007; Bartlett et al., 2008), syllables are, technically speaking, phonological entities that can only be composed of strings of phonemes. Most linguists view syllables as an important unit of prosody because many phonological rules and con- straints apply within syllables or at syllable bound- aries (Blevins, 1995). Apart from their purely linguistic significance, syllables play an important role in speech synthesis and recognition (Kiraz and M¨obius, 1998; Pearson et al., 2000). The pronunciation of a given phoneme tends to vary depending on its location within a syl- lable. While actual implementations vary, text-to- speech (TTS) systems must have, at minimum, three components (Damper, 2001): a letter-to-phoneme (L2P) module, a prosody module, and a synthesis module. Syllabification can play a role in all three modules. Because of the productive nature of language, a dictionary look-up process for syllabification is in- adequate. No dictionary can ever contain all possi- ble words in a language. For this reason, it is neces- sary to develop systems that can automatically syl- labify out-of-dictionary words. In this paper, we advance the state-of-the-art in both categorical (non-statistical) and supervised syllabification. We outline three categorical ap- proaches based on common linguistic theories of syllabification. We demonstrate that when imple- mented carefully, such approaches can be very ef- fective, approaching supervised performance. We also present a data-driven, discriminative solution: a Support Vector Machine Hidden Markov Model (SVM-HMM), which tags each phoneme with its syllabic role. Given enough data, the SVM-HMM achieves impressive accuracy thanks to its ability to capture context-dependent generalizations, while also memorizing inevitable exceptions. Our ex- periments on English, Dutch and German demon- strate that our SVM-HMM approach substantially outperforms the existing state-of-the-art learning ap- proaches. Although direct comparisons are difficult, our system achieves over 99% word accuracy on German and Dutch, and the highest reported accu- racy on English. The paper is organized as …

Content truncated to 500 words