According to the Shpinxtrain docs one might be better off training word models rather than phone models when having a small vocabulary, so; I made a full word phone list, matched each word to itself in the dictionary, made sure all words were in the training material and trained.
I works (but only as good as my phone model)... but I guess I will have better results if I tweak the train / decode config parameters to match a word model rather then a phone model!?
However, I have trouble finding information on how to work with a word model?
Anyone else using this setup?
Thanks
Johan
BTW, downloaded, built, setup, trained and tested the Sphinx3 package on XP, MS-VS2005, ActivePerl, Swedish ms-wavs in less than two days, all with only minor code, script changes.
- Brilliant work guys!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
According to the Shpinxtrain docs one might be better off training word models rather than phone models when having a small vocabulary, so; I made a full word phone list, matched each word to itself in the dictionary, made sure all words were in the training material and trained.
I works (but only as good as my phone model)... but I guess I will have better results if I tweak the train / decode config parameters to match a word model rather then a phone model!?
However, I have trouble finding information on how to work with a word model?
Anyone else using this setup?
Thanks
Johan
BTW, downloaded, built, setup, trained and tested the Sphinx3 package on XP, MS-VS2005, ActivePerl, Swedish ms-wavs in less than two days, all with only minor code, script changes.
- Brilliant work guys!
Hi Johan,
Are you using more than one phone per model in your dictionary? Example; instead of:
word word
use:
word word1 word2 word3
if the word would normally be modeled by 3 phones.
I believe I found this approach on the Sphinx-site or in the forum.
Best Regards
Morten