Menu

Help with pocketsphinx

Help
puffadder
2009-10-29
2012-09-22
  • puffadder

    puffadder - 2009-10-29

    Hi all, I am working with pocketsphinx. I am having an ultra-small vocabulary
    (20-30 words) and my end goal is word spotting. I don't know whether my
    questions have been asked (or more importantly, answered) by someone but I
    couldn't find them anywhere so I am posting the same:

    ** I am able to train with word models wherein the dictionay maps each word to itself and the phone file contains the list of words. But when I try to run pocketsphinx_continuous, it gives an error of unknown phone(s). It refuses to regard words as phones. This problem was not there when I used sphinx3 decoder. Is it so that pocketsphinx does not support word models or am I doing something wrong ??? Here is the exact format of error:

    ERROR: "dict.c", line 556: 'THE': Unknown phone 'THE'
    ERROR: "dict.c", line 440: Failed to add THE to dictionary
    ERROR: "dict.c", line 556: 'THERE': Unknown phone 'THERE'
    ERROR: "dict.c", line 440: Failed to add THERE to dictionary

    ** Regarding non-speech events, I have added non-speech events to the filler dictionary and I want them to be recognized as filler words
    instead of words from the grammar. Do I need train with these non-speech
    events (e.g. tap, breath, clap etc.) or is the support for recognizing fillers
    inbuilt in pocketsphinx??? Also, -fillprob isn't helping
    my cause, whenever such a non-speech event occurs, it is recognized as a
    speech word with a high confidence score (i.e. pprob).

    Here is my filler dictionary.....

    ++BADRECORDING++ +GARBAGE+

    ++BEEP++ +NOISE+

    ++CHAIRSQUEAK++ +NOISE+

    ++CLICK++ +SMACK+

    ++COUGH++ +NOISE+

    ++CROSSTALK++ +GARBAGE+

    ++DISKNOISE++ +NOISE+

    ++DOOROPEN++ +NOISE+

    ++DOORSLAM++ +NOISE+

    ++EXHALATION++ +BREATH+

    ++FOOTSTEPS++ +NOISE+

    ++KNOCKING++ +NOISE+

    ++LAUGHING++ +NOISE+

    ++LAUGHTER++ +NOISE+

    ++LIPSMACK++ +SMACK+

    ++LOUDBREATH++ +BREATH+

    ++MICROPHONEMVT++ +NOISE+

    ++MISCNOISE++ +NOISE+

    ++MOVEMENT++ +NOISE+

    ++OTHERMOUTHSOUND++ +NOISE+

    ++PAPERRUSTLE++ +NOISE+

    ++PHONERING++ +NOISE+

    ++POORMICPOSITION++ +GARBAGE+

    ++SIGH++ +BREATH+

    ++SINGING++ +GARBAGE+

    ++SLOW++ +GARBAGE+

    ++SNIFF++ +BREATH+

    ++SOFT++ +GARBAGE+

    ++SQUEAK++ +NOISE+

    ++TAP++ +NOISE+

    ++THROATCLEAR++ +BREATH+

    ++THUMP++ +NOISE+

    ++TONGUECLICK++ +SMACK+

    ++TYPING++ +NOISE+

    ++UNINTELLIGIBLE++ +GARBAGE+

    Thanks in advance.

     
  • Nickolay V. Shmyrev

    > Is it so that pocketsphinx does not support word models or am I doing
    something wrong ???

    You are doing something wrong. Most likely you are using wrong acoustic model
    which has no phones from your dictionary. You should learn to provide full log
    instead of tiny excerpt.

    Another thing you are definitely doing wrong is that you are using the single
    phone and thus same amount of states per words of different length. Basically
    it's not an optimal approach to a model training.

    > Do I need train with these non-speech events (e.g. tap, breath, clap
    etc.) or is the support for recognizing fillers inbuilt in pocketsphinx???

    Model should have such phones trained.

     
  • puffadder

    puffadder - 2009-10-30

    Thanks for the prompt response.

    -- Another thing you are definitely doing wrong is that you are using the
    single phone and thus same amount of states per words of different
    length. Basically it's not an optimal approach to a model training.

    What is the remedy? I mean how can I make the states per words vary
    proportional to the length of my words?

     
  • puffadder

    puffadder - 2009-10-30

    Edit: Regarding non-speech events, what is the modus operandi to train
    fillers?

    I am at a loss to figure out the correct way of doing this and this is hurting
    my recognition accuracy badly.

     

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.