hypseg param not working in sphinx3_align

[Pranav Jawale]

Hello,

I built sphinx3 from SVN, (downloaded on 9th March). I am using sphinx3_align
with -hypseg parameter

INFO: cmd_ln.c(512): Parsing command line:
..\sphinx3_align \
    -hmm E:\users\pranav\sphinx\sphinxtest\test1\hmm1 \
    -dict E:\users\pranav\sphinx\sphinxtest\test1\lm1\test1.dic.txt \
    -fdict E:\users\pranav\sphinx\sphinxtest\test1\lm1\test1.filler.txt \
    -hyp E:\users\pranav\sphinx\sphinxtest\test1\hypseg.out.txt \
    -hypseg E:\users\pranav\sphinx\sphinxtest\test1\hypseg_align.txt \
    -cepdir E:\users\pranav\sphinx\sphinxtest\test1\feats \
    -cepext .mfc \
    -ceplen 13 \
    -frate 100 \
    -ctl E:\users\pranav\sphinx\sphinxtest\test1\list_one \
    -debug 3 \
    -insert_sil 0 \
    -insent E:\users\pranav\sphinx\sphinxtest\test1\insent.txt \
    -outsent E:\users\pranav\sphinx\sphinxtest\test1\outsent.txt \
    -phlabdir E:\users\pranav\sphinx\sphinxtest\test1 \
    -stsegdir E:\users\pranav\sphinx\sphinxtest\test1 \
    -s2stsegdir E:\users\pranav\sphinx\sphinxtest\test1 \
    -wdsegdir E:\users\pranav\sphinx\sphinxtest\test1 \
    -phsegdir E:\users\pranav\sphinx\sphinxtest\test1

But the hypseg file is never created. Also in the sourcecode I could not find
section where writing hypseg job is done (like the one in sphinx3_decode). Is
this param working fine with anyone?

Thanks.

[Pranav Jawale]

Actually I need unscaled acoustic scores, (that's why I was going for hypseg)
and I'm not sure whether the scores given in wdseg files have been unscaled or
not (there is no -hypsegscore_unscale option in sphinx3_align).

[Nickolay V. Shmyrev]

Hello

wdseg is unscaled.

[Pranav Jawale]

Thanks.

[Pranav Jawale]

Hello,

I did one test. I compared the decode acoustic scores with align scores. When
I keep -hypsegscore_unscale no only then the decoder score is near to the
align score

Decoder ASCR for a word (with -hypsegscore_unscale yes) = 993077
Decoder ASCR for the same word (with -hypsegscore_unscale no) = -902992
Align score from wdseg = -913451
I added following code to write_stseg function

        mdef_phone_str(kbc->mdef, stseg->pid, str2);

        fprintf(fp, "\nFrameIndex %d Phone %s PhoneID %d SenoneID %d state %d Ascr %11d \n",i,str2,stseg->pid,stseg->sen,stseg->state,stseg->score);

str2 was defined as char str2; by me

Some example lines which were created in stseg file were -

FrameIndex 0 Phone SIL PhoneID 18 SenoneID 54 state 0 Ascr           0

FrameIndex 1 Phone SIL PhoneID 18 SenoneID 55 state 1 Ascr      -29907

FrameIndex 2 Phone SIL PhoneID 18 SenoneID 56 state 2 Ascr      -35769

FrameIndex 3 Phone a PhoneID 19 SenoneID 57 state 0 Ascr      -12950

FrameIndex 4 Phone a PhoneID 19 SenoneID 57 state 0 Ascr      -16943

FrameIndex 5 Phone a PhoneID 19 SenoneID 57 state 0 Ascr      -18500

FrameIndex 6 Phone a PhoneID 19 SenoneID 57 state 0 Ascr       -9099

And in wdseg I have

     SFrm  EFrm    SegAScr Word
        0     2     -65676 <s>
        3    91    -913451 ananasa
       92    96     -91163 </s>
 Total score:    -1070290

The sum of acoustic scores of states of first 3 frames ( 0 -29907 -35769) is
exactly equal to wdseg of . Same thing happens for the second word.

As all the state level scores are negative, I think scaling by the best senone
is not undone anywhere in the code.

These were the best senone scores for the first 3 frames-

NumFrame 97
0 28088
1 -24821
2 -36011

[Pranav Jawale]

Update: It seems there are two kinds of scalings used in sphinx3_align

1. scaling senone scores with best senone scores in the frames
2. scaling by 'bestscore' in a frame - in align_frame() relevant code is

     s->newscore -= bestscore;   /* Scale, to avoid underflow */
   

Only second kind of scaling is undone in build_stseg() . The code is

  stseg->score = h->score - prevscr + score_scale[f];

In order to make the scores compatiable with decoder unscaled scores further
unscaling will be reqd I guess.

[Pranav Jawale]

It may be possible to unscale wrt best senones by keeping track of
ascr->cache_best_list and using it when the stseg scores are updated in
backtrace (build_stseg, build_phseg, build_wdseg etc.). I tried it but it gave
some other error (somehow stseg->pid was assigned a very high value, assertion
failed at one place). Anyway, it can be fixed I guess if somebody needs ..