next up previous contents index
Next: Speaker selection Up: Speech recognition systems Previous: Rejection mode


Application vocabulary and confusion matrix  


The technology provider performs different evaluation tests. These can be performed with laboratory  databases as well as with field  databases (speech recorded during exploitation). The confusion matrix permits a more detailed evaluation of the system capabilities than the usual recognition rate, with respect to a given database. It gives the confusion between words used to evaluate the performance. An example with some isolated French words is given in Table 2.1 for illustration [Choukri et al. (1988)]:


Classer Placer Annexer
Classer 66 22 5
Placer 19 59 1
Annexer 9 7 76
Table 2.1: Confusion matrix 

This matrix shows that in 22 cases (out of 93) the word ``Classer'' is recognised as ``Placer''. This matrix is very important because it can be taken into account to design an accurate recogniser  through an adequate vocabulary  selection where the most confusable series are withdrawn and replaced by synonyms or equivalent terms. It can also be considered at the application level to provide an optional candidate not supplied by the recogniser .


EAGLES SWLG SoftEdition, May 1997. Get the book...