pocketsphinx confidence more than 100%

[Yuri Orlov]

Hi,
I use the next code to get posterior probability for recognition result:
int32 prob = ps_get_prob(recognizer->decoder, &uttid);
The value what I got for prob is 1213251
Than I use the next code to get confidence:
double64 conf = logmath_log_to_ln(ps_get_logmath(recognizer->decoder), prob);
The value what I got for conf is 121.34
Is there something wrong, or this mean a bug by posterior probability
calculation or by confidence calculation?

[Nickolay V. Shmyrev]

Can you please provide more information on this issue - audio files, decoder
logs, models.

[Yuri Orlov]

I have got this problem using unimrcp and pocketsphinx plugin. I have tried
reproducing it with the pocketsphinx_batch, but instead I get always
confidence 0. I have used the same audio, model and pocketsphinx version.
This is the link with data : http://www.file-
upload.net/download-2822667/de8KHz.zip.html
I have modified pocketsphinx_batch code by adding the next stuff:
prob = ps_get_prob(ps, &uttid);
conf = logmath_log_to_ln(ps_get_logmath(ps), prob);
E_INFO("Posterior probability: %d; confidence: %.7f\n", prob);
This log message can be found in de8KHz-1-1.log
Regards,
Yuri

[Nickolay V. Shmyrev]

Hm, I succeded to reproduce this problem. It looks like a regression because
it works ok in older version. Let me investigate this issue.

[Nickolay V. Shmyrev]

Thanks, this bug was fixed in trunk. Btw, to print posterior use this code

               int32 prob = ps_get_prob(ps, &uttid);
                float conf = logmath_exp(ps_get_logmath(ps), prob);
                E_INFO("Posterior probability: %d; confidence: %g\n", prob, conf);

Your code has few mistakes (conf is missing, wrong function is used for log).

[Yuri Orlov]

Thanks!