How does mk_mllr_work?

[morten_aau]

Hi all,

Does anyone know how the tool for creating regression classes (mk_mllr_class) when doing speaker adaptation works?

Best regards
Morten

[The Grand Janitor]

mk_mllr_class -example yes

-a

[morten_aau]

Hi,

I tried "mk_mllr_class -example yes" and the output is shown here:

---

mr@posh4:[adaptation]$ ./bin/mk_mllr_class -example yes
./bin/mk_mllr_class \ -example yes

Example:
mk_mllr_class -nmap mapfile -nclass 4 -cb2mllrfn out.cb2mllr

INFO: parse_cmd_ln.c(127): User asked for help or example.

I could not make this work, and I the tried the following (a hint to this was given at: http://www-2.cs.cmu.edu/~archan/presentation/Adaptation20041104.pdf):

./bin/mk_mllr_class -nmap 544 -nclass 12 -cb2mllrfn out.cb2mllr < phone_classes/approximinants_senone.lst

where "544" is the number of lines in "phone_classes/approximinants_senone.lst" which is a list of senone numbers (id's) taken from the model definition file (only senone id's belonging to the class I want to adapt are included).

I tried to use the binary file that was output when running mllr_solve, but I get the following error:

FATAL_ERROR: "s3cb2mllr_io.c", line 85: out.cb2mllr cb2mllr[0] -> 192 which is >= n_mllr_class (12)

I guess I'm using the tools wrong?

Best regards
Morten