Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Comparative Performance in Large-Vocabulary Isolated-Word Recognition in Five European Languages

James Barnett, Paul Bamberg, Martin Held, Juan Huerta, Linda Manganaro, Adam Weiss

Dragon Systems, Inc., Newton, MA, USA

Dragon Systems' DragonDictateŽ for Windows Version 1.0 (DDWin) large-vocabulary isolated-word dictation system is available in several languages. This paper reports on comparative recognition performance in five of them: French (pre-release), German, Italian, Spanish, and English. The tests are based on a variety of different document styles, including a work of philosophy, a popular novel, and a software user's manual. The results show statistically significant differences between the languages, with Italian being the easiest to recognize, and German and French the hardest.
An error analysis shows the importance of vocabulary coverage in inflected languages, as English shows an out-of-vocabulary error rate of less than .5%, while the other four languages have rates above 2%. French showed a much higher rate of homophone errors than the other languages. However, recognition performance is uniformly better than that reported in a similar test in 1991.

Full Paper

Bibliographic reference.  Barnett, James / Bamberg, Paul / Held, Martin / Huerta, Juan / Manganaro, Linda / Weiss, Adam (1995): "Comparative performance in large-vocabulary isolated-word recognition in five european languages", In EUROSPEECH-1995, 189-192.