question about expected WER for an4 tutorial

[goose]

Good morning,

I just ran the an4 tutorial from http://cmusphinx.sourceforge.net/wiki/tutori
alam and the final decoder
output is

SENTENCE ERROR: 76.2% (99/130) WORD ERROR RATE: 33.9% (261/773)

I am to new to it to know if this is as expected. Out of curiosity i used the
training data as the test data

cp an4_train.fileids an4_test.fileids
cp an4_train.transcription an4_test.transcription

and got

SENTENCE ERROR: 37.0% (351/948) WORD ERROR RATE: 13.6% (739/5442)

I was wondering if these values are in the regular range or if something is
wrong with my setup. I would suggest/request that the tutorial gives some
indication which values to be expected for the an4 test case (I assume it cant
be given exactly) so that new users can use these values as an indication if
everything works correct.

Thanks a lot

martin

(i am running this on Mac Lion)

[Nickolay V. Shmyrev]

If you didn't edit N_TIED_STATES, this result is expected. If you put
N_TIED_STATES to 200, accuracy must be around 15-18% on test data.

I would suggest/request that the tutorial gives some indication which values
to be expected for the an4 test case

This is indicated in tutorial already:

SENTENCE ERROR: 70.8% (92/130) WORD ERROR RATE: 30.3% (233/773)

[goose]

thanks a lot, I didnt make the link between the given WER values in the
tutorial and the actual data which was stupid as it should have been clear
from the number of words/sentences