.ptm. / .semi. on Android

[dohceit Fota]

i trained an acoustic model on SphinxTrain using cont. parameter with the
following spec.s :

---

corpus description
dictionary 1800 words
training hours 45
audio 16KH wav format

---

then i decoded using S3Decoder and got good results as follows:

wer 4%
accurcy 97%

---

then converted the the audio to 8khz using 'sox' command
and trained two models the first .ptm. and the other is .semi. , using the
following config :(see the link)

the semi AM tested on PocketSphinx_PATCH on my linux box (fedora14), using
same training material for testing, and got excellent result, but no good
result on different speaker (speaker not used on training)

then put the .semi. on android with the following feat. parameters:

-alpha 0.97
-doublebw no
-nfilt 31
-ncep 13
-lowerf 200
-upperf 3500
-nfft 512
-wlen 0.0256
-transform legacy
-feat s2_4x

---

and the following configuration in the android app using an4.cd_semi_1000

-hmm /sdcard/edu.cmu.pocketsphinx/hmm/en_US/hub4wsj_sc_8k/
-backtrace yes
-beam 1e-48
-topn 4
-lm 10

---

unfortunetly i got horruble result !!

how can i improve the results ..

Thanks

[hiyassat]

Dear Nickolay
i face the same problem and i did exactly the same, my configuration as
follows
-alpha 0.97
-doublebw no
-nfilt 31
-ncep 13
-lowerf 200
-upperf 3500
-nfft 512
-wlen 0.0256
-transform legacy
-feat s2_4x

-backtrace yes
-beam 1e-48
-topn 4
-lm 6.5

but accuracy is very bad, the system is very fast , but it does recognize any
correct words, all the output words are wrong

can you help please ?

[Nickolay V. Shmyrev]

Try to increase default beams. Make sure that during training feature
extraction tool used proper sample rate 8000.

[hiyassat]

thank you Nickolay, i would like to tell you that pocketSphinx now working
excellent ,when i did feature exctraction i forget to change the sample rate
to 8000