Generation and Selection of Pronunciation Variants for a Flexible Word Recognizer

This paper presents an approach for the generation and selection of pronunciation transcriptions for a exible word recognizer. The basic idea is to produce pronunciation variants and corresponding scores with a set of pronunciation variation rules, which are weighted with their frequencies of occurence measured on the training data. This approach addresses the problem of interfering transcriptions of diierent words producing recognition errors. As an extreme test case, pronunciation variants are produced by segmenting the graphemic string and replacing the graphemic segments into phonemic ones, according to a grapheme-phoneme cluster pair alphabet that covers all possible grapheme-phoneme relations in German. Results show that even with such inaccurate \pronunciation variation" rules recognition is at least as good as with standard pho-netic transcriptions, without even a need for a lexicon. This is of course a major advantage for a exible word recognizer.