sphinx-3 tuning tool

[Berker Batur]

Hi,

Is there any tool or script that automatize tuning in Sphinx-3 ?
If not, what are the common steps, configuration variables to change, etc.. we
should do
when tuning the decoder.

Thanks.

[Nickolay V. Shmyrev]

Hi, Is there any tool or script that automatize tuning in Sphinx-3

1 2 3 4 5 6 7 8 9 10 11 12

for lw in 11.5
do
    # '1e-1' '1e-15' '1e-40' '1e-60' '1e-80' '1e-100' '1e-120' '1e-140' '1e-180' '1e-300' '1e-1000'
    for beam in '1e-50'
    do
        # '1e-1' '1e-15' '1e-40' '1e-60' '1e-80' '1e-100' '1e-140' '1e-300'
        for wbeam in '1e-50'
        do
            # 0.01 0.1 0.2 0.3 0.4 0.5 0.6 0.7 1 3
            for wip in 0.2
            do
                echo "lw:$lw beam:$beam wbeam:$wbeam wip:$wip num_jobs:$num_jobs result:$result list_file:$list                perl scripts/decode_ps.pl $lw $beam $wbeam $wip $num_jobs $result $list_file;
            done
        done
    done
done

If not, what are the common steps, configuration variables to change, etc..
we should do when tuning the decoder

http://cmusphinx.sourceforge.net/wiki/decodertuning

[Berker Batur]

Thanks a lot.

[Berker Batur]

Hi,

The script that you wrote, will run in exponential with number of each
configuration variable.
Because, it will look at all configuration possibilities.
For ex. : If one decoding takes 0.3 hours, the script below will run nearly
100 days.

I guess, we should use some searching algorithms in tuning (like Hill
Climbing) for discarding irrelevant states
and decreasing the run time.

Berker

[Nickolay V. Shmyrev]

The script that you wrote, will run in exponential with number of each
configuration variable.

Yes, the goal was to give you and idea

I guess, we should use some searching algorithms in tuning (like Hill
Climbing) for discarding irrelevant states

That would be very interesting to have this as a part of CMUSphinx training
process! But remember that accuracy function is not convex, so many methods
don't really apply as is.

[Berker Batur]

That would be very interesting to have this as a part of CMUSphinx training
process! But remember that accuracy function is not convex, so many methods
don't really apply as is.

I did not understand what you mean exactly by "accuracy function is not
convex". Could you give any paper written about that,
or any other source that mention about this problem would be good.

I am trying to code a script for optimization of sphinx3 decoder accuracy.

Berker.

[Nickolay V. Shmyrev]

I did not understand what you mean exactly by "accuracy function is not
convex"

When function is not convex many local minima exists and most optimization
methods will only find local minimum, not a global one. Probably you'll be
interested in introduction to numeric optimization like

http://books.google.ru/books?id=epc5fX0lqRIC&lpg=PP1&dq=numerical%20optimizat
ion&pg=PP1#v=onepage&q&f=false

or

http://books.google.ru/books?id=88X5vN3ybvgC&lpg=PR1&dq=numerical%20optimizat
ion&pg=PR1#v=onepage&q&f=false