Sixth European Conference on Speech Communication and Technology
Two methods of assessing the acoustic modelling of an automatic speech recognition system are presented. The first is objective, and is based on determination of the phone substitution matrix. For this, a generalized alignment procedure is introduced, which leads to better results. The second method is subjective. Hypothesis words, as segmented by the recognizer, are evaluated by test subjects on their validity. Without additional context, humans accept about half of the false alarms found by a word spotter.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Leeuwen, David A. van / Louwere, Michael de (1999): "Objective and subjective evaluation of the acoustic models of a continuous speech recognition system", In EUROSPEECH'99, 1915-1918.