Confidence Score

[Anonymous]

How is the score for the corresponding hyp calculated? It seems to be changed
in ps_search_hyp from ps_end_utt(). After this it becomes really unclear what
the math behind it is.

I ask because for my command and control application, I would like to set
thresholds for whether or not to accept the command based on the mean and
standard deviation. This score seems to vary a lot leading to a large
acceptance range. Knowing how the score is calculated, I believe will help me
find a more reliable way to set the thresholds.

Would you recommend setting these thresholds a different way?

Thanks in advance!

[Nickolay V. Shmyrev]

The score returned in ps_get_hyp is the score of the path, it indeed has no
meaning for you. To get the confidence score you need the following function

/** 
 * Get posterior probability. 
 * 
 * @note Unless the -bestpath option is enabled, this function will 
 * always return zero (corresponding to a posterior probability of 
 * 1.0).  Even if -bestpath is enabled, it will also return zero when 
 * called on a partial result.  Ongoing research into effective 
 * confidence annotation for partial hypotheses may result in these 
 * restrictions being lifted in future versions. 
 * 
 * @param ps Decoder. 
 * @param out_uttid Output: utterance ID for this utterance. 
 * @return Posterior probability of the best hypothesis. 
 */
POCKETSPHINX_EXPORT
int32 ps_get_prob(ps_decoder_t *ps, char const **out_uttid);

Be careful it requires some specific configuration. Also remember that
posterior probability as a confidence measure requires you to use a diverse
grammar in order to cover background speech. It might be garbage model or just
grammar that is big enough.

[Anonymous]

Can a garbage model be created by adding words other than those in the
commands to the filler dictionary?

[Anonymous]

Nevermind about that last question. Is it known what causes a partial result?