Morphological Decomposition for Asr in German

In this contribution we report on our ongoing work in lexical decomposition for automatic speech recognition (ASR). Lexical decomposition is investigated with a twofold goal: lexical coverage optimization and improved automatic letter-tosound conversion. Whereas morphological decomposition is a widely-studied domain in linguistics, our interest is limited here to identifying and processing the statistically most relevant sources of lexical variation in text corpora. Lexical variation is shown to be particularly important for nouns, due to compounding. A set of about 340 decomposition rules has been developed using statistics from 300M words from different newspaper sources (primarily 14 years from the TAZ, the Berliner TAgesZeitung). The out-of-vocabulary (OOV) rate on the same 300M words is reduced from 5.2 to 4.6% in case-sensitive form and to 4.2% in case-insensitive form. For letter-to-sound conversion cross-morpheme letter sequences are a major source of ambiguity. Decomposition, by reducing these ambiguities, contributes to producing more consistent phonemic transcriptions for pronunciation dictionaries.