Weighted finite-state transducer approach to German compound words reconstruction for Speech Recognition

Abstract

An approach is proposed for German Large Vocabulary Speech Recognition, dealing with the problem of compound words, based on unsupervised word decomposition for German words and a probabilistic method for combining the words using finite state transducers. The basic idea of the method is to train n-gram language model on the texts where compound words are substituted by their parts plus concatenation symbol. Thus, the context information is taken into account for the compound words and is used in the process of recombination to find most probable variant for recognition result. The advantage of this approach is the improvement of the word recognition accuracy and a more precise recombination of compound words.