Say-as classification for alphabetic words in Japanese texts

Modern Japanese texts often include Western sourced words written in Roman alphabet. For example, a shopping directory in a web portal, which lists more than 8,000 shops, includes a total of 6,400 alphabetic words. As most of them are very new and idiosyncratic proper nouns, it is impractical to assume all those alphabetic words can be registered in the word dictionary of a text-to-speech synthesis system; their pronunciations must be derived automatically. Our solution consists of two steps. Step 1 classifies each unknown alphabetic word into a say-as class (English, Japanese, French, Italian or English spell-out), which indicates how it is to be read, and Step 2 derives the pronunciation using the grapheme-to-phoneme conversion rules for the classified sayas class. This paper proposes a method of say-as classification (i.e. Step 1) that uses the Support Vector Machine. After some trial and error, we achieved 89.2% accuracy for web shop data, which we think sufficient for practical use.