Speech Recognition Engines

Flat
Julius error, triphone not found -- even through triphone is present in training corpus
User: Olumide
Date: 7/27/2013 8:32 pm
Views: 5034
Rating: 3

I'm getting several dozen of the following type of errors while experimenting with forced alignment with julius, using the voxforce acoustic model

Error: voca_load_htkdict: line 40: triphone "aa-f+ah" not found
Error: voca_load_htkdict: the line content was: 33  [PROPHECY]     p r aa f ah s iy

Error: voca_load_htkdict: the line content was: 33  [PROPHECY]     p r aa f ah s iy

Error: voca_load_htkdict: line 45: triphone "ae-dh+ah" not found

Error: voca_load_htkdict: line 199: triphone "aa-l+ah" not found
Error: voca_load_htkdict: line 199: triphone "l-ah+jh" not found
Error: voca_load_htkdict: the line content was: 169 [KNOWLEDGE]    n aa l ah jh

For a while I though these errors occur because the training corpus did not include such words (knowledge and prophecy for example), but it turned out that these words are in the training corpus. As such I find these errors very difficult to explain. Why aren't these triphones in the HMM (hmmdefs) even though they are in the training corpus?

Thanks.

 

--- (Edited on 7/27/2013 8:34 pm [GMT-0500] by Olumide) ---

Re: Julius error, triphone not found -- even through triphone is present in training corpus
User: azeem
Date: 8/7/2013 9:25 am
Views: 1894
Rating: 2

There may be two reasons behind that problem:

 

1) maybe you are trying to do Forced Alignment with the words "knowledge" and "prophecy" with a different pronunciation with regards to the one that has been given during training.

For example, in the Voxforge dictionary that I have downloaded, "prophecy" is spelled like this:

PROPHECY        [PROPHECY]      p r aa f ax s iy

which is different from "[PROPHECY]     p r aa f ah s iy" that is posted in your post above.

Also, "knowledge" is spelled in Voxforge like these

KNOWLEDGE       [KNOWLEDGE]     n aa l ax jh

KNOWLEDGE(2)    [KNOWLEDGE]  n aa l ix jh

and your pronunciation is different again: [KNOWLEDGE]    n aa l ah jh

I think that if you're using your pronunciations with Voxforge acoustic models, it's probable that those triphones are unseen, because even if those words are present in the training set, they may have a different pronunciation.

 

2) make sure that all the phones that you provide for forced alignment are present in the acoustic model (i.e., in the hmmdefs file). ..but I guess you did this already :-)

--- (Edited on 8/7/2013 9:25 am [GMT-0500] by azeem) ---

PreviousNext