The problem is that I am given empty matrices for some speakers. By looking at the MLF file the common thing about them is that the utterances for these speakers are too short. Setting the value of HADAPT:SPLITTHRESH to zero doesn't help either. Any idea where the problem is?
The conclusion was that for some utterancs the number of frames is just so small that HTK puts out empty matrices. Its safe to leave out such utterances while doing CMLLR transform estimation.
I am using HERest for generating CMLLR matrices. The config file for this purpose is as follow:
ReplyDeleteHADAPT:TRANSKIND = CMLLR
HADAPT:USEBIAS = TRUE
HADAPT:REGTREE = regtree.tree
HADAPT:ADAPTKIND = TREE
HADAPT:SPLITTHRESH = 1000.0
HADAPT:BLOCKSIZE = "IntVec 3 13 13 13"
The problem is that I am given empty matrices for some speakers. By looking at the MLF file the common thing about them is that the utterances for these speakers are too short. Setting the value of HADAPT:SPLITTHRESH to zero doesn't help either. Any idea where the problem is?
The conclusion was that for some utterancs the number of frames is just so small that HTK puts out empty matrices. Its safe to leave out such utterances while doing CMLLR transform estimation.
ReplyDelete