[hts-users:02322] Bugs related to CMLLR, SAT, SEMIT, and MLLRCOV

[Junichi Yamagishi <jyamagis@xxxxxxxxxxxx>]

Hi all,

John Dines from IDIAP found that CMLLR transforms estimated by
HTK and HTS version of HERest are completely different from one another
even for the normal HMMs (not HSMM case).

The HTS version of HERest seems to frequently output messages like

 Not updating transform (Block: 1 Row: 1 Iter: 1 (26.671926 28.735340))
 Not updating transform (Block: 1 Row: 3 Iter: 1 (10.560113 10.623189))

or sometimes warnings like

 WARNING [-1]  WARNING: accumulates incorrect (-36690.361990 < 0) -  
resetting in
 WARNING [-1]  WARNING: accumulates incorrect (-41041.051185 < 0) -  
resetting in

In fact WERs using the CMLLRs transforms caused huge differences on  
our data:
 HTK HEREst:1.3%   HTS HERest:12.4%


Anyway, recently, Heiga and I worked on this issue and found *two*  
relatively
critical bugs of HAdapt.c. I report the first bug in this mail and Heiga
would report the second bug later.

The first one was included in the InvertG function of HAdapt.c

/* InvertG: invert matrix G (band or full structure) */
static void InvertG (DMatrix G, DMatrix invG, const int c, const int  
bsize, const int bandw, const Boolean uBias)
{
  DMatrix inv, u, v;
  DVector w;
  int i,j,cnti,cntj,st,en,dim,size;

-  if (bandw<bsize) {
+  if (bandw+1<bsize) {
     /* matrix size */
    for (st=c-bandw; st<1; st++);
     for (en=c+bandw; en>bsize; en--);


The original one was simply wrong. This affects not only
adaptation demos bug also semi-tied covariance estimation of
speaker-dependent demos. I hence recommend everyone to fix this soon.

Regards,
Dr. Junichi Yamagishi
University of Edinburgh






--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.