VoxForge
VoxForge
Hi!
I'm trying to test your acoustic model, but it seems that the tiedlist is missing a lot of triphones, which causes Julius to exit. Is it normal? Is there a way of dealing with that, other than adding a line in the tiedlist for each missing phone and linking them to actual triphones, as said in this topic, and is not always trivial?
I'm testing on grammar with words which are supposed to have been used for the training, so there shouldn't be any problem. But one thing that I remarked is that they always correspond to multiple pronunciations of a word, for example:
Error: voca_load_htkdict: line 12: triphone "l-ix+p" not found
Error: voca_load_htkdict: the line content was: 6 [philip(2)] f ih l ix p
Error: voca_load_htkdict: line 7: triphone "*-hh+w" or biphone "hh+w" not found
Error: voca_load_htkdict: line 7: triphone "hh-w+eh" not found
Error: voca_load_htkdict: the line content was: 3 [when(2)] hh w eh n
Error: voca_load_htkdict: line 9: triphone "*-hh+w" or biphone "hh+w" not found
Error: voca_load_htkdict: line 9: triphone "hh-w+ih" not found
Error: voca_load_htkdict: the line content was: 3 [when(4)] hh w ih n
Did you train the model with multiple pronunciation?
Thanks in advance,
Marion
Hi Marion,
>But one thing that I remarked is that they always correspond to multiple
>pronunciations of a word, for example
This might have something to do with Step 8 - Realigning the Training Data, which says:
[...] the HVite command can consider all pronunciations for each word (in the case where a word has more than one pronunciations), and then output the pronunciation that best matches the acoustic data.
Try changing your sample.voca file to use the first pronunciation of a word rather than one of the alternatives.
It may be that we need to train the VoxForge acoustic model with more speech containing the pronunciation variation you are looking for.
Ken