VoxForge
Hi Ken,
I try to run hebrew Acoustic Model follow by the tutorial since step 2 (as you already recommended . . .) to step 10.
I tried to append the follow new Hebrew words (As copied from dict file).
ARBA [ARBA] aa r b ax sp
EFES [EFES] ih f eh s sp
EKHAD [EKHAD] ih hh ax d sp
EYAL [EYAL] eh y ax l sp
HALAKHUT [HALAHUT] hh ae l ae hh uh t sp
HATEMPERATURA [HATEMPERATURA] hh ae t eh m p eh r ae t uh r ae sp
KHAMESH [KHAMESH] hh ae m eh sh sp
KHAYEG [KHAYEG] hh ae y eh g sp
LIAV [LIAV] l ih ae v sp
MA [MA] m aa sp
SENT-END [] sil
SENT-START [] sil
SHALOSH [SHALOSH] sh ae l ao sh sp
SHESH [SHESH] sh eh sh sp
SHEVA [SHEVA] sh eh v ae sp
SHMONE [SHMONE] sh m ao n eh sp
SHTAYM [SHTAYM] sh t ae y ih m sp
SIGI [SIGI] s ih g ih sp
STAV [STAV] s t ae v sp
TESHA [TESHA] t eh sh ae h sp
YAHAV [YAHAV] y ae hh ae v sp
til step 10 it looked just fine.
But . . . I stuched at step 10 when I run the command, I quote:
HHEd -A -D -T 1 -H hmm12/macros -H hmm12/hmmdefs -M hmm13 tree.hed triphones1
As you can see, the stucking point is "ERROR [+2662] FindProtoModel: no proto for k in hSet"
TB 350.00 ST_h_4_ {}
Tree based clustering
Start h[4] : 1 have LogL=-49.533 occ=16.7
Via h[4] : 1 gives LogL=-49.533 occ=16.7
End h[4] : 1 gives LogL=-49.533 occ=16.7
TB: Stats 1->1 [100.0%] { 261->72 [27.6%] total }
TR 1
Adjusting trace level
AU fulllist
Creating HMMset using trees to add unseen triphones
ERROR [+2662] FindProtoModel: no proto for k in hSet
FATAL ERROR - Terminating program HHEd
Could you please help me to get over the error?
HHEd -A -D -T 1 -H hmm12/macros -H hmm12/hmmdefs -M hmm13 tree.hed triphones1
Thanks,
Eyal
Hi Eyal,
Step 10 requires the creation of a tree.hed script which contains language specific "Questions". I've never had to create my own tree.hed script.
For more information on how to create a tree.hed file for a new language, see the following links:
See also the HTK manual.
You should be able to use the monophone acoustic models you created in Step 8 (located hmm9) of the tutorial, and using monophones1 rather than tiedlist. Note: you likely won't get as good recognition as with triphone acoustic models.
Ken