With the help of the sphinx library, a program for recognizing Russian speech is implemented. Phonemes are drawn from the dictionary with phonemes words.JSGF and word forms proper words.dic. The question arose how to derive the percentage of word recognition? For example, after the cessation of recognition, by clicking on the button, the percentage of recognition was determined. As is customary, everywhere they say that the percentage of recognition is 90. How to achieve this in the program. Program implementation on c #

    2 answers 2

    I'm not sure about:

    1. You give to the program what you are going to say in text form.

    2. You pronounce this pre-prepared phrase you compare the result of recognition with the help of the LCS algorithm with what you previously transmitted.

    3. You get% similarity.

    Only in this way, in my opinion, it is possible to get% recognition, if the library itself cannot natively give this information.

    I also found info on the off site:

    http://cmusphinx.sourceforge.net/wiki/tutorialtuning

    It looks like what I described.

    • I will try. Thanks. If something happens or fails, accomplish your goal. - Ivan Antonyuk
    • LCS algorithm. Do not quite understand. The fact is that the phrase can be any. I didn't understand the algorithm a bit. Is there an algorithm code? - Ivan Antonyuk February
    • @ ivanantonyuk, I refer to added. On the off site there is a description. - iluxa1810
    • But the example is described in the event that it is sound files, models so to speak. JSpeech Grammar Format (JSGF) dictionary is connected to me - Ivan Antonyuk

    Could not "stick" the command, which is listed on the offsite. Make sure to add -samprate 8000 to the above command if you are decoding 8kHz files! It would work if you worked on it.

    The point is that I use the dictionary and phonemes, and not the static model word-align.pl.