Towards Automatic Transcription of Spontaneous Czech Speech in the MALACH Project

Our paper discusses the progress achieved during one-year effort with building the Czech LVCSR system for an automatic transcription of spontaneously pronounced testimonies in the MALACH project [1]. The difficulty of this task stems from the highly inflectional nature of the Czech language and is further multiplied by the presence of many colloquial words in spontaneous Czech speech and also by the need to handle emotional speech filled with disfluencies, heavy accents, age-related coarticulation and language switching. In this paper we concetrate mainly on the acoustic issues - the proper choice of the front-end parameterization, handling the non-speech events in acoustic modeling and especially the unsupervised usage of the MLLR adaptation technique. A method for selecting suitable language model data is also briefly mentioned.