This paper addresses the problem of robust multilingual speaker-independent connected digit recognition over telephone lines. We test and compare various techniques for signal conditioning, improved model representation and model training across three different languages, namely Danish, English and Spanish, and evaluate their language dependency. Word errors rates between 0.5% and 1.0% have been achieved for all languages on various speech databases collected over the public telephone network.
Bibliographic reference. Salavedra, Joan / Jacobsen, Claus / Rahim, Mazin G. / Zeljkovic, Ilija / Wilpon, Jay G. (1995): "Multi-lingual connected digits recognition", In EUROSPEECH-1995, 2119-2122.