Modeling Reduced Pronunciations in German

This paper deals with pronunciation modeling for automatic speech recognition in German with a special focus on reduced pronunciations. Starting with our 65k full form pronunciation dictionary we have experimented with different phone sets for pronunciation modeling. For each phone set, different lexica have been derived using mapping rules for unstressed syllables, where schwa-vowel+[l n m] are replaced by syllabic [l n m]. The different pronunciation dictionaries are used both for acoustic model training and during recognition. Speech corpora consisted of television programmes, which contain signal segments of a varying acoustic and linguistic nature. The speech is produced by a wide variety of speakers, with linguistic styles ranging from prepared to spontaneous speech and with changing background and channel conditions. Experiments were carried out using 4 news programmes and documentaries lasting more than 15 minutes each (with a total of 1h20min). Word error rates obtained vary between 19 and 29%, depending on the programme and the system configuration. Only small differences in recognition rates were measured for the different experimental setups, with slightly better results obtained by the reduced lexica. 146 Adda-Decker & Lamel