Sphinx 3 force-alignment problem for few words

[nayan kalita]

Hi Everyone,

I am having an issue of wrong alignment and low phone score on Sphinx3 force-alignment for few sets of words. and the alignment is fine for all others words excepts these words.

The list of such words are Father(F AA DH ER), Polish(P AA L IH SH),Swiss(S W IH S),Miss(M IH1 S) etc.
The major problem i am facing for word "Father", the alignment is incorrect and score of F is low for all occurance. I am giving a example below.(file name : 91_kash_androidOctober_1477365933528_91) .

 SFrm  EFrm   SegAScr Phone
    0    80  -3574250 SIL
   81   101   -173969 SIL
  102   104 -247371495 F SIL AA b
  105   107   -262382 AA F DH i
  108   110   -224049 DH AA ER i
  111   113   -140673 ER DH SIL e
  114   130    -65163 SIL

Total score: -251811981

The pronouncation for word Father ( F AA DH ER) seems to be okay for me.

For your reffernce i have uploaded the few wave files and the acoustic mode used on the below link.

https://drive.google.com/drive/folders/0B00sD6w9Ioe9VFNtOC1rWlR5XzQ?usp=sharing

Please help me on this to improve the alignment.

[Nickolay V. Shmyrev]

You need to provide the command line, sphinx3 and sphinxbase version, the dictionary and all other data files to reproduce your problem.

[nayan kalita]

Hi Nickolay,

Thanks for your reply. I am using sphinx3 version 3.08 and sphinxbase version 0.6.

I am using sphinx_fe with following Parameter to genarate the mfcc file.

sphinx_fe -i Data/91_kash_androidOctober_1477365933528_91.wav -argfile etc/feat.params -ei wav -mswav yes -eo mfc -o feat/91_kash_androidOctober_1477365933528_91.mfc

and then run the following command for force-alignment.

sphinx3_align -hmm models/ -dict etc/voxeduEnglish.dic -fdict etc/words.filler -ctl temp/91_kash_androidOctober_1477365933528_91.ctl -insent temp/91_kash_androidOctober_1477365933528_91.sent -cepdir feat -phsegdir temp -phlabdir temp -wdsegdir temp -outsent 91_kash_androidOctober_1477365933528_91.out

The detail log report for command line display is given in the attached .zip file . File name is Log.txt

[Nickolay V. Shmyrev]

Same command with default en-us model aligns fine, so it should be the issue with your acoustic model not trained appropriately.

[nayan kalita]

Hi Nickolay,

Thank you for your reply.

1) Please suggest us the best acoustic model for US english for sphinx3.

2) If any open source databases are available to train a acoustic model

[Nickolay V. Shmyrev]

1) Please suggest us the best acoustic model for US english for sphinx3.

default models available in downloads are ok

2) If any open source databases are available to train a acoustic model

tedlium, librispeech

[nayan kalita]

Hi Nickolay,

Thank you for your suggestions.

1) default models available in downloads.Here you are referring to latest cmu acoustic model cmusphinx-en-us-5.2 .

2) I am able to find out the download link for librispeech database. http://www.openslr.org/12/ . Please share the download page link for tedlium database.

[Nickolay V. Shmyrev]

1) default models available in downloads.Here you are referring to latest cmu acoustic model cmusphinx-en-us-5.2 .

Yes

2) I am able to find out the download link for librispeech database. http://www.openslr.org/12/ . Please share the download page link for tedlium database.

http://www.openslr.org/19/

[nayan kalita]

Hi Nickolay,

Thank you for your help.