VoxForge
At the point to create hmm11:
HERest -A -D -T 1 -C config -I wintri.mlf -t 250.0 150.0 3000.0 -S train.scp -H hmm10/macros -H hmm10/hmmdefs -M hmm11 triphones1
It give me many warning like this:
WARNING [-2331] UpdateModels: t+ae[1] copied: only 1 egs
in HERest
WARNING [-2331] UpdateModels: t+ay[3] copied: only 2 egs
in HERest
....
It's a problem, or it's all ok?
Because I reach the end of tutorial to create my personal acoustic model, but when I try to use it with Julian it give me some errors:
Reading in dictionary...
line 3: triphone "*-f+ow" or biphone "f+ow" not found
line 3: triphone "f-ow+n" not found
> 2 [PHONE] f ow n
line 4: triphone "*-k+ao" or biphone "k+ao" not found
line 4: triphone "k-ao+l" not found
> 2 [CALL] k ao l
line 5: triphone "d-ay+ax" not found
> 3 [DIAL] d ay ax l
line 6: triphone "t-iy+v" not found
> 4 [STEVE] s t iy v
line 8: triphone "b-aa+b" not found
> 4 [BOB] b aa b
line 9: triphone "*-jh+aa" or biphone "jh+aa" not found
line 9: triphone "jh-aa+n" not found
line 9: triphone "aa-n+s" not found
> 4 [JOHNSTON] jh aa n s t ax n
line 10: triphone "*-jh+aa" or biphone "jh+aa" not found
line 10: triphone "jh-aa+n" not found
> 4 [JOHN] jh aa n
line 11: triphone "*-jh+ao" or biphone "jh+ao" not found
line 11: triphone "jh-ao+r" not found
line 11: triphone "r-d+ax" not found
> 4 [JORDAN] jh ao r d ax n
line 13: triphone "f-ay+v" not found
> 5 [FIVE] f ay v
line 15: triphone "n-ay+n" not found
> 5 [NINE] n ay n
line 21: triphone "th-r+iy" not found
> 5 [THREE] th r iy
line 23: triphone "z-ih+r" not found
line 23: triphone "ih-r+ow" not found
> 5 [ZERO] z ih r ow
////// Missing phones:
*-f+ow or biphone f+ow
*-jh+aa or biphone jh+aa
*-jh+ao or biphone jh+ao
*-k+ao or biphone k+ao
aa-n+s
b-aa+b
d-ay+ax
f-ay+v
f-ow+n
ih-r+ow
jh-aa+n
jh-ao+r
k-ao+l
n-ay+n
r-d+ax
t-iy+v
th-r+iy
z-ih+r
//////////////////////
error in reading sample.dict: 12 words failed out of 23 words
ERROR: failed to read dictionary, terminated
If I try to use monophones model It start but all the recognizes are wrong
Tks
Manuel
Hi Manuel,
I think you have two problems:
> WARNING [-2331] UpdateModels: t+ae[1] copied: only 1
egs
in HERest
This is basically HTK telling you have don't enough audio data for certain triphones. Your acoustic model will compile, but this will affect your recognition rates. So if you want to improve recognition, you need to add more audio data to your acoustic model. You should try to add words that contain the triphones that are missing (i.e. record a sentence containing a word with the "t+ae" triphone in it, etc.), especially if that word is in your grammar file.
>line 3: triphone "*-f+ow" or biphone "f+ow"
not found
>line 3: triphone "f-ow+n" not found
>>
2 [PHONE] f ow n
This error is telling you that the phones in your grammar file don't match the triphones (series of 3 phones) in your acoustic model (i.e. hmmdef file). Your grammar pronunciations need to match the pronunciations the in your pronunciation dictionary (which you used to compile your acoustic model). I think you're using one pronunciation dictionary for training and another for your grammar pronunciations - they need to be the same.
Hope that helps,
Ken
>> WARNING [-2331] UpdateModels: t+ae[1] copied: only 1
egs
in HERest
>This is basically HTK telling you have don't enough audio data for >certain triphones. Your acoustic model will compile, but this will >affect your recognition rates. So if you want to improve >recognition, you need to add more audio data to your acoustic >model. You should try to add words that contain the triphones >that are missing (i.e. record a sentence containing a word with >the "t+ae" triphone in it, etc.), especially if that word is in your >grammar file.
How many words I have to add to solve this kind of Warning? For example
WARNING [-2331] UpdateModels: aw-t[9] copied: only 1 egs
in HERest
I added following word within this triphone: ALLENTOWN FREETOWN, GEORGETOWN, OUT, OUTBOARD. But warning message is the same.
why?
Tks
Manuel
Hi Manuel,
>> WARNING [-2331] UpdateModels: t+ae[1] copied: only 1
egs
in HERest
From the HTK book (all the error codes are at the back of the HTK book):
−2331 Not enough training examples
Model was not updated as there were not enough training examples. Either reduce the minimum speci?ed by -m or use more data.
>How many words I have to add to solve this kind of Warning?
usually 3-5 words (if I remember correctly to get HTK to compile the acoustic model - you still need lots speech more to improve recognition rates) with the triphone suffice... you just picked some words that do not contain "aw-t" - so you really only have two words with "aw-t":I just used trial-and-error for alot of this type of thing ... just add more words until the warning goes away and use that as your rule of thumb for other triphones.
Ken