lattice-tool: HTK to PFSG

james.waterhouse at elf.mcgill.ca james.waterhouse at elf.mcgill.ca
Fri Sep 30 13:05:46 PDT 2005


Andreas,

Thanks for the response to my question about lattice-tool and fsm and pfsg
lattices.

I have another question.... Could you perhaps tell me how the PFSG scores are
calculated from the HTK scores? I thought it was straight forward...

weight = amscale*a + lmscale*l + word_penalty

but when I looked at the PFSG output for a simple example it was obviously not
that simple. I am supposed to look at the code and try and figure things out
but if you can tell me in a line or two what is happening then it would be
greatly apreciated.

Thanks,
James

p.s. here is the simple example and the result...

VERSION=1.0
UTTERANCE=/opt/asrtools/apps/parya/ttimit/test/dr1/faks0/si1573.mfc
lmname=/opt/asrtools/apps/parya/ttimit/wdnet
lmscale=8.10   wdpenalty=-9.40
vocab=/opt/asrtools/apps/parya/ttimit/dict
N=6 L=5
I=0    t=0.00  W=!NULL
I=1    t=0.02  W=sent-start          v=1
I=2    t=0.04  W=sil4                v=1
I=3    t=0.09  W=sil4                v=1
I=4    t=0.12  W=t                   v=1
I=5    t=0.15  W=s                   v=1
J=0     S=0    E=1    a=-249.59   l=0.000
J=1     S=1    E=2    a=-204.42   l=0.000
J=2     S=2    E=3    a=-354.02   l=-6.250
J=3     S=3    E=4    a=-218.17   l=-3.900
J=4     S=4    E=5    a=-215.62   l=-2.440

Resulting PFSG....

name /opt/asrtools/apps/parya/ttimit/test/dr1/faks0/si1573.mfc
nodes 11 NULL NULL NULL NULL NULL NULL sent-start sil4 sil4 t s
initial 0
final 5
transitions 10
0 6 -2712479
1 7 -2260756
2 8 -4263106
3 9 -2714179
4 10 -2570411
6 1 0
7 2 0
8 3 0
9 4 0
10 5 0



More information about the SRILM-User mailing list