3

I found that there are so many parameters to tune up speech recognition( for example input parameters in pocketsphinx_continuous app). Many parameters are enumerations, other are floating values. Where can I find the range of variation for each parameter for pocketsphinx_continuous app?

BrettWatts
  • 87
  • 9

1 Answers1

1

Most of pocketsphinx parameters are internal for the algorithms used in speech recognition, they are not trivially connected and it is not really useful to change them unless you understand what is going on.

A single parameter change might require change in other parameters or might break the recognition in very rare but significant cases.

Generally parameters are almost optimal or close to that. You can vary every parameter slightly (+/- 10% relative) within the current value if you are interested, but a significant change is rarely reasonable thing.

If you are interested in some decoding tuning, it is better to describe what goal do you want to archive.

Nikolay Shmyrev
  • 24,897
  • 5
  • 43
  • 87
  • 1
    I need a recognition API that can work with little dictionary (50-100 words). I have only 2 language models for my native language (Russian) for CMU Sphinx. It is a big problem to find good recognition API for russian c++ developers. I think I need to tune recognintion. Because when I speak to microphone a word that is NOT in my dictionary, grammar, transcription pocketsphinx_continuous detects this as a word FROM my dictionary. My dictionary has only 2 words. – BrettWatts Mar 17 '15 at 17:53
  • 1
    This behavior is expected. To search for the words in continuous stream you can use keyword spotting mode, see http://stackoverflow.com/questions/28918032/raspberry-pi-asynchronous-continuous-speech-recognition-in-python/28918542#28918542. You can specify the list of keyphrases and detection threshold for them. If you want some command system you can define a single activation keypharse and switch to grammar search to recognize keywords. – Nikolay Shmyrev Mar 17 '15 at 18:33