How can you adapt the model without adding the unseen models
required for adaptation data?
First of all, duration model format changed from HTS-2.1.
If you use old duration models built using HTS 2.0 or before,
please change the duration model to new format like this.
e.g. 5-state HSMMs
cat <<EOF >! output.2.1/convert
SU 5 1 1 1 1 1
EOF
HHEd -A -B -C general.conf -T 1 -D -V -H output/clustered.dur.mmf -
w output.2.1/clustered.dur.mmf -i -p output.2.1/convert output/
context.dur.list
Then, add unseen models first
HHEd -A -B -C general.conf -T 1 -D -V -H clustered.cmp.mmf -w
clustered.cmp.mmf -i -p unseen.cmp context.cmp.list
HHEd -A -B -C general.conf -T 1 -D -V -H clustered.dur.mmf -w
clustered.dur.mmf -i -p unseen.dur context.dur.list
where unseen.cmp includes
LT tree.mcep.inf
LT tree.logF0.inf
AU context.cmp.list.train+adapt
and unseen.dur includes
LT tree.dur.inf
AU context.dur.list.train+adapt
Then adapt the models using MAP like this
HERest -A -B -C general.conf -C map.conf -D -V -H
clustered.cmp.mmf -N clustered.dur.mmf -I adapt.mlf -M output/ -R
output -S ADAPTDATA -T 1 -h \"*/%%%*.cmp\" -m 0 -t 5000 5000 30000 -
u mvwpdmvp -w 3 context.cmp.list context.dur.list
where map.conf includes
HMAP:MAPTAU = 50
HMAP:DURMAPTAU = 50
HMAP:MIXWEIGHTFLOOR = 3
and general.conf includes
APPLYVFLOOR = T
NATURALREADORDER = T
NATURALWRITEORDER = T
VFLOORSCALESTR = "Vector 5 0.01 0.01 0.01 0.01 0.01"
DURVARFLOORPERCENTILE = 1.000000
APPLYDURVARFLOOR = T
MAXSTDDEVCOEF = 10
Please follow these procedures.
If the problem happens again, please let me know again.
I'll check source codes.
Junichi
On 2008/03/28, at 5:53, tu qixiong wrote:
Hi Dr.Junichi,
I still have a question with adaptation of duration. After doing
HERest with the setting "HERest -u mvwpdmvp -w 3", I found just
the files clustered_cmp.mmf and fullcontext_dur.mmf had been
modified, but the file clustered_dur.mmf stayed the same, So the
duration wouldn't be adapted.
In order to add duration information of adaption data to the
original models,I tried doing HHed to modify the file
clustered_dur.mmf, but the file trees_dur.inf would be modified at
the same time, and couldn't be used when test. I don't know how to
solve this problem.
PS:I used HTS2.1RC1, thank you.
On 3/26/08, Junichi Yamagishi <jyamagis@xxxxxxxxxxxx> wrote:
On 2008/03/26, at 9:52, tu qixiong wrote:
> But when I do HERest with the spectification "HERest -u mvwpdmvp -w
> 3", the system says:
>
> ERROR [+2320] SetuFlags: Unknown update flag m.
>
> What's the matter? I use HTS2.0.
First of all, please use HTS-2.1RC1.
> And can I view -u mvwp as adaptation of cmp and -u dmvp as
> adaptation of duration?
Yes.
> Is the adaptation of cmp and duration completed at the same time
> When doing -u mvwpdmvp?
Yes, it is, if you use HTS-2.1RC1.
--
yours, sincerely
QiXiong Tu