Ney H.: \Recent Improvements of the RWTH Large Vocabulary Speech Recognition System on Spontaneous Speech

fsixtus,molau,kanthak,s hlueter,neyginformatik.rwth-aa hen.de This paper presents re ent improvements of the RWTH large vo abulary ontinuous spee h re ognition system (LVCSR). In parti ular, we will report on the integration of a ross-word models into the rst re ognition pass, and des ribe better algorithms for fast vo al tra t normaliza-tion (VTN). We will fo us both on the improvements in word error rate and how to speed up the re ognizer with only minimal loss in re ognition a ura y. Implementation details and experimental results are given for the VerbMo-bil task, a German spontaneous spee h orpus. The 25.0% word error rate (WER) of our within-word baseline system was redu ed to 21.4 % with VTN and a ross-word models. De reasing the real-time fa tor (RTF) by up to 85 % resulted in only a small degradation in re ognition performan e of 2 % relative on average. 1