Menu

Some questions

Help
Eugene
2016-12-05
2016-12-05
  • Eugene

    Eugene - 2016-12-05

    Hi. Got couple of questions.
    1. Is this correct that I can improve quality of recognition if I remove all words in dictionary that I don't need to recognize, and leave only 100 words for example that I need to recognize?
    2. If 1 is true, is it true then that to improve recognition of 'no', for example, I should add 'go' to dictionary, so it can differ them. Though I don't need to recognize 'go' by itself.

     

    Last edit: Eugene 2016-12-05
    • Arseniy Gorin

      Arseniy Gorin - 2016-12-05
      1. Is this correct that I can improve quality of recognition if I remove all words in dictionary that I don't need to recognize, and leave only 100 words for example that I need to recognize?

      No, dictionary is for training. You can improve by narrowing the language model

      1. If 1 is true, is it true then that to improve recognition of 'no', for example, I should add 'go' to dictionary, so it can differ them.

      It depends if you care more about precision or recall

       
  • Eugene

    Eugene - 2016-12-05

    Hm. I thought dictionary is where possible results are stored. I left just 'no N OW' and 'yes Y EH S' in it, and these two words are recognized 100% of times. When I use full dictionary I get different results for saying 'yes' and 'no'.

     
    • Arseniy Gorin

      Arseniy Gorin - 2016-12-05

      Well it depends what you actually run.
      Technically the decoding hypothesis is determined by language model. Its likely that your code builds language model from the dictionary.

       
  • Eugene

    Eugene - 2016-12-05

    I run pocketsphinx, -hmm -lm -dict that are used, are from model folder with english model that is included with source code.

     
    • Arseniy Gorin

      Arseniy Gorin - 2016-12-05

      In general using large LM and limiting dictionary is not a right approach.
      I would suggest to rather look at grammar.

      About the accuracy, it will depend on what your speakers say. If you have 'yes/no' grammar and speaker says 'bo', sphinx will likely recognize 'no'. On the other hand, the larger grammar you have, the more there is a chance that correctly pronounced 'no' will be recognized as soething else.

      You can also check keyword spotting mode to tune precision/recall with word threshold

       
  • Eugene

    Eugene - 2016-12-05

    Tested, if I add 'go G OW' to dictionary that has 'no' and 'yes' already, it doesn't mistake 'go' for 'no' after that, when I say 'go'.

    What do you mean by looking at grammar?

     
  • Eugene

    Eugene - 2016-12-05

    Ok, thanks!

     

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.