Pocketsphinx - accuracy issues

  • Gaurav Kumar

    Gaurav Kumar - 2014-08-19

    I’m using Pocketsphinx in my app for Speech Recognition purpose with a very small vocabulary of 2 - 4 words only in Language Model. But I'm getting very unexpected results using this. In my Language Model, I have set "JOHN". Now when I say "JOHN" it recognises as JOHN which is correct. Now when I say "JOE", "JUG", "JOLLY", "JAPAN" it still recognises it as "JOHN".

    Means we are getting the same results i.e. "JOHN" is recognised when I say "JOE", "JUG", "JOLLY", "JAPAN".

    I'm testing on real device iPhone 4S iOS version 7.1.2. Can you please explain this & How can I get better accuracy in word recognition.

    Looking forward to an early reply. Let us know if you need any info from our end.

  • Nickolay V. Shmyrev

    If you want to recognize words like JOE or JUG you need to add them to the grammar. If you want to cover all possible words you can use keyword spotting mode from latest pocketsphinx or you can include standalone phonemes into the grammar to model the other words. It's hard to make such system accurate though since it's very hard to distinguish JAN from JOHN.

  • Gaurav Kumar

    Gaurav Kumar - 2014-08-20

    Dear Nickolay,

    Thanks for the reply.

    However, we want to recognize only JOHN and not JOE, JUG etc.

    Pl. note user is free to say whatever he wants. So, if we have setup recognition only for JOHN, in that case, if the user says JOE, JUG, JOLLY etc. then they should not get recognized as JOHN.

    Hope my requirement is clear to you.

    I understand that JAN and JOHN are close. However JUG and JOHN are not close at all?

    Can you pl. guide more on this.

    • Nickolay V. Shmyrev


      You need to provide decoder all possible alternatives including the words you want ignore in the future. If you do not want to name all these words you can include a loop of phones into your grammar for them, then you recognizer will know that it can expect either word JOHN or loop of phones.

      If you want to look for only a set of keyphrases, you can check recently added keyword model which essentially adds phone loop automatically. You can try how it works by looking for keyword "oh mighty computer" in android demo



Log in to post a comment.

Get latest updates about Open Source Projects, Conferences and News.

Sign up for the SourceForge newsletter:

JavaScript is required for this form.

No, thanks