After intensive research real-time, speaker-independent, large vocabulary recognition of continuos speech has become a field where a variety of commercial products is available meanwhile. This success opened the perspective to also cope with the recognition of several different languages; this problem will become of vital interest and commercial potential in the future, for example, in multilingual regions like Europe, in the tourist business, or simply due to often quoted 'globalization' of markets and societies.
The thesis is among the first to adress the problem of multilingual automatic speech recognitionbg . It investigates the approaches of porting, cross-language recognition, and simultaneous multilingual recognition and it deals with dialects and the problems of non-native speakers. Expermiments are carried out using speech samples from seven languages German, English, Italian, Czech, Slovak, Slovenian and Japanese. The variety of languages, compare recognition rates using different approaches, investigate the influence of language models, and to suggest promising strategies for multilingual speech recognition - although there seems to be no general best strategy for all subsets of languages so far. Multilingual speech recognition (real-time, speaker-independent, large vocabulary, continuous speech) will remain a challenge for some time. But this work demonstrates that it is a feasible problem and provides many interesting findings to solve it.fg v