next up previous
Next: Results Up: The IDIAP system Previous: Classifier

Threshold settings

In order to decide if a test speech segment was said by the target speaker, an a priori decision threshold has to be set. The threshold tex2html_wrap_inline184 chosen here is derived from the Furui threshold setting method [1, 2].

tex2html_wrap_inline186

tsp=target speaker, ntsp=non-target speaker

An extended threshold determination is used here:

tex2html_wrap_inline192

in this case, the followed transformation is applied:

tex2html_wrap_inline194

so the threshold tex2html_wrap_inline196 becomes speaker independent, and it becomes possible to adjust the threshold to improve the cost function (see gif). The data used as non-target speaker data (for threshold setting) came from the training set of the 1996 NIST evaluation data. In order to determine tex2html_wrap_inline198 and tex2html_wrap_inline200 the non-target speaker data were "passed through" each target speaker model to obtain tex2html_wrap_inline198, tex2html_wrap_inline200 and the three constants A,B,C.



Dominique Genoud
Mon Aug 18 15:56:59 MET DST 1997