Joint Processing and Discriminative Training for Letter-to-Phoneme Conversion

Proceedings of ACL-08: HLT, pages 905–913, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Joint Processing and Discriminative Training for Letter-to-Phoneme Conversion Sittichai Jiampojamarn† Colin Cherry‡ Grzegorz Kondrak† †Department of Computing Science ‡Microsoft Research University of Alberta One Microsoft Way Edmonton, AB, T6G 2E8, Canada Redmond, WA, 98052 {sj,kondrak}@cs.ualberta.ca colinc@microsoft.com Abstract We present a discriminative structure- prediction model for the letter-to-phoneme task, a crucial step in text-to-speech process- ing. Our method encompasses three tasks that have been previously handled separately: input segmentation, phoneme prediction, and sequence modeling. The key idea is online discriminative training, which updates parameters according to a comparison of the current system output to the desired output, allowing us to train all of our components together. By folding the three steps of a pipeline approach into a unified dynamic programming framework, we are able to achieve substantial performance gains. Our results surpass the current state-of-the-art on six publicly available data sets representing four different languages. 1 Introduction Letter-to-phoneme (L2P) conversion is the task of predicting the pronunciation of a word, repre- sented as a sequence of phonemes, from its or- thographic form, represented as a sequence of let- ters. The L2P task plays a crucial role in speech synthesis systems (Schroeter et al., 2002), and is an important part of other applications, including spelling correction (Toutanova and Moore, 2001) and speech-to-speech machine translation (Engel- brecht and Schultz, 2005). Converting a word into its phoneme represen- tation is not a trivial task. Dictionary-based ap- proaches cannot achieve this goal reliably, due to unseen words and proper names. Furthermore, the construction of even a modestly-sized pronunciation dictionary requires substantial human effort for each new language. Effective rule-based approaches can be designed for some languages such as Spanish. However, Kominek and Black (2006) show that in languages with a less transparent relationship be- tween spelling and pronunciation, such as English, Dutch, or German, the number of letter-to-sound rules grows almost linearly with the lexicon size. Therefore, most recent work in this area has focused on machine-learning approaches. In this paper, we present a joint framework for letter-to-phoneme conversion, powered by online discriminative training. By updating our model pa- rameters online, considering only the current system output and its feature representation, we are able to not only incorporate overlapping features, but also to use the same learning framework with increasingly complex search techniques. We investigate two on- line updates: averaged perceptron and Margin In- fused Relaxed Algorithm (MIRA). We evaluate our system on L2P data sets covering English, French, Dutch and German. In all cases, our system outper- forms the current state of the art, reducing the best observed error rate by as much as 46%. 2 Previous work Letter-to-phoneme conversion is a complex task, for which a number of diverse solutions have been pro- posed. It is a structure prediction task; both the input and output are structured, consisting of sequences of letters and phonemes, respectively. This makes L2P a poor fit for many machine-learning techniques that are …

Content truncated to 500 words