German regional variants - a problem for automatic speech recognition?

A well known problem in automatic speech recognition (ASR) is robustness against the variability of speech between speakers. There are several ways to normalise different speakers; one of them is to deal with the problem of regional variation. In this paper we discuss the problem of whether moderate regional variants of German in uence the automatic speech recognition process and whether there is a way to improve performance through knowledge of the regional origin of the unknown speaker. The basic idea in our experiment is to cluster test speakers into distinct dialectal regions and derive observations about the typical pronunciation within these regions from a classi ed training set. In a cheating experiment where the origin of the test speakers is known we verify whether the use of the dialect-speci c pronunciation forms will improve the overall performance of the recognizer. It turns out that simply using dialect-speci c pronunciation does not signi cantly improve word accuracy on the VERBMOBIL 1996 task.