Noisedict in pocketsphinx

[Mark]

My breath is being picked up as a "misunderstanding" instead of filtered out and ignored. I checked the noisedict files, in location \hmm\wsj1\ the file ontains:

<s> SIL
</s> SIL
<sil> SIL
++NOISE++ +NOISE+
++BREATH++ +BREATH+
++SMACK++ +SMACK+

In locations \model\communicator\ and \libs\Communicator_semi_40.cd_semi_6000\ has the following in their noisedict files:

++AE++ AE
++AH++ AA
++BACKGROUND++ +BACKGROUND+
++BREATH++ +BREATH+
++COUGH++ +COUGH+
++EH++ EH
++ER++ ER
++LAUGH++ +LAUGH+
++MM++ M
++MUMBLE++ +BACKGROUND+
++NOISE++ +NOISE+
++OH++ OW
++SMACK++ +SMACK+
++UH++ +UH+
++UH_NOISE++ +UH+
++UM++ +UM+
++UM_NOISE++ +UM+

I see ++BREATH++ +BREATH+ in all of them but still my breath is being picked up. Why?

[Mark]

Not all the settings just this setting "fillprob"

[Nickolay V. Shmyrev]

During search for the answer decoder tries to insert fillers after each word. The insertion is penalized with probability so you can tune the frequency of fillers. So with a large filler probability decoder will choose filler more likely than word from the grammer. With smaller probability it will prefer grammar words.

[Nickolay V. Shmyrev]

Well, probably your breath is not the real breath. To get the idea you need to find out how breath sounds in the training dabase (wsj, which is unfortunately not availble for free). Probably if you'll ask David will give you examples. Also you can try to play with filler probability -fillprob. The bigger values will give fillers more preference.

[Mark]

First, who is David and how do I get a hold of him?

Second, how would I set -fillprob, on windows xp if that matters.

Thanks.

[Nickolay V. Shmyrev]

David is the author of pocketsphinx. See this post

https://sourceforge.net/forum/message.php?msg_id=5748449

About -fillprob, in freeswitch find the following lines

    ps->config = cmd_ln_init(ps->config, ps_args(), FALSE,
                                                     "-samprate", rate,
                                                     "-hmm", model,
                                                     "-lm", lm,
                                                     "-dict", dic,

add "-fillprob", "0.1"

there

[Mark]

I tried adding "-fillprob", "0.1",
but that seemed to make phrase recognition worst.

What do all these settings mean?