I am implementing an ASR System for Spanish language using Sphinx-4. I
implemented the frontend, the language model, the dictionary and the acoustic
model. I tested it with sphinx-3 (perl scripts_pl/decode/slave.pl) and i got
WER=5.4%. After, i tested my Sphinx-4 system (i changed the tidigits
performance test to adapt it to my files) and i got WER=21.8%. I used the same
test data for the test, and i expected similar WER but they are very
different.
I uploaded the test folder, the .wav and etc files at the following link:
See the difference between featureExtraction vs featureExtractor. Once you'll
fix the name accuracy will be way better. I will fix the sphinx4 so such error
will be way easier to detect. I wanted to do that long time ago.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2010-08-12
Hello Nickolay,
I can solve the problem with your help, I really appreciate it. Thank you very
much.
Fidelia
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Good evening,
I am implementing an ASR System for Spanish language using Sphinx-4. I
implemented the frontend, the language model, the dictionary and the acoustic
model. I tested it with sphinx-3 (perl scripts_pl/decode/slave.pl) and i got
WER=5.4%. After, i tested my Sphinx-4 system (i changed the tidigits
performance test to adapt it to my files) and i got WER=21.8%. I used the same
test data for the test, and i expected similar WER but they are very
different.
I uploaded the test folder, the .wav and etc files at the following link:
http://rapidshare.com/files/411114016/Consulta.rar
Thank you very much.
Fidelia
Hello Fidelia
After some poking I finally found the reason:
Your pipeline was:
But later
See the difference between featureExtraction vs featureExtractor. Once you'll
fix the name accuracy will be way better. I will fix the sphinx4 so such error
will be way easier to detect. I wanted to do that long time ago.
Hello Nickolay,
I can solve the problem with your help, I really appreciate it. Thank you very
much.
Fidelia