CMU Sphinx / Forums / Help: PocketSphinx Accuracy

PocketSphinx Accuracy

Forum: Help

Creator: Senjam Shantirani

Created: 2016-04-25

Updated: 2016-05-13

Senjam Shantirani - 2016-04-25

Following the Acouctic Model creation example using an4 I got the accuracy of 28.1% , which is almost same as given in the website.

Now, I got accuracy of 61% or WER of 39% using TIMIT.
Is this accuracy proper with TIMIT, based on any past study?

TIMIT is weird but for my current study I am using it, as I am struggling to get better data.

Regards,
Senjam Shantirani

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
- Nickolay V. Shmyrev - 2016-04-25
  
  It depends on parameters and langauge model you used.
  
  You can download librispeech database instead of timit.
  
  If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Senjam Shantirani - 2016-05-13

They come in .flac.
If I convert them to 16 KHz Mono and .wav using ffmpeg, will they give a good result..
Also, I guess it is free corpus.
Also is WSJ a free corpus? If not, can you suggest me some free corpus besides Librespeech.

Please advice..

Last edit: Senjam Shantirani 2016-05-13

If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
- Nickolay V. Shmyrev - 2016-05-13
  
  If I convert them to 16 KHz Mono and .wav using ffmpeg, will they give a good result..
  
  Yes
  
  Also is WSJ a free corpus?
  
  No, wsj is not free
  
  If not, can you suggest me some free corpus besides Librespeech.
  
  tedlium
  
  Last edit: Nickolay V. Shmyrev 2016-05-13
  
  If you would like to refer to this comment somewhere else in this project, copy and paste the following link:

Log in to post a comment.