Open alexanderkoller opened 5 years ago
I found the following in the source: (0 map anything to grammar, 100 reject if it's not a perfect match)
and:
thresholds must be converted to be useful in Sphinx.
TIMO tested this with a bunch of files containing digits
except for one or two and tested a grammar that only allowed
those two digits. The range of false positives varied within
a range of 10e-25 and 10e-50 (anything accepted as one|two at 10e-50).
thus 0 -> 10e-50
and 1 -> 10e-25
If no value is set, the default is 0 (i.e., also accept low confidence values)
there are two possibilities in the speech recognition: (a) follow the paths in the grammar, (b) follow an out-of-grammar path. The probability of (b) is set with the confidence threshold.
I.e., if you set a high threshold, it's likely that the speech recognizer follows (b) and the result is that you get out-of-grammar except if what you speak sounds very much like what the grammar expected. Equivalently, if you set a low threshold, the speech recognizer is very unlikely to follow (b) and anything you say will sound like one of the possibilities in the grammar.
I don't know where this documentation should go, so I'm leaving it here for the moment.
By default, the confidence textfield in the ASR node properties is blank. This makes it hard for a user to guess what kind of value they should enter here. Can we initialize the contents of the textfield with the default confidence value?