Menu

Difference between pocketsphinx_continuous and pocketsphinx_batch

Help
tfpeach
2015-10-19
2015-11-09
  • tfpeach

    tfpeach - 2015-10-19

    Hi, dear all,

    I trained a acoustic model and decode the testing data with pocketsphinx_batch, the WER is about 2.5%. I want to get the timing boundary of the recognized word, however, it seems there is no such function in pocketsphinx_batch. So I turned to pocketsphinx_continous. But the recognition seems different from the one given by pocketsphinx_batch (even with the same parameters). I then changed a little bit. I use the recognized text from pocketsphinx_batch as the grammar and do the decoding with pocketsphinx_continuous, some of the files say there are errors during decoding and cannot give out the result.

    I want ask:
    1. what's the difference between these two decoder?
    2. how can I get timing boundary with pockesphinx_batch? (or say is there anyway to get timing boundary)

    Thank you.

     
    • Nickolay V. Shmyrev

      1. what's the difference between these two decoder?

      Batch uses LiveCMN, continuous uses continuous CMN which is not always accurate from initial estimation.

      1. how can I get timing boundary with pockesphinx_batch? (or say is there anyway to get timing boundary)

      -hypseg option creates file with word times.

       
      • tfpeach

        tfpeach - 2015-10-23

        Thank you.

        I notice that the output is always a little shifted from our label. Is that because of the feature extraction machenism in Sphinx? As I know, some toolkits does not start from the first singal point, because they need a little bit information ahead to compute the feature. Could you tell me where I can find the details about the feature extraction in Sphinx? I tried but didn't find them.

        Thank you.

         
  • tfpeach

    tfpeach - 2015-10-21

    Does anyone else have the same problem as mine?

     
  • Nickolay V. Shmyrev

    I notice that the output is always a little shifted from our label. Is that because of the feature extraction machenism in Sphinx? As I know, some toolkits does not start from the first singal point, because they need a little bit information ahead to compute the feature.

    Maybe

    Could you tell me where I can find the details about the feature extraction in Sphinx? I tried but didn't find them.

    In source code

     

Log in to post a comment.