Acoustic Scores and its relation to frames and confidence

[bassem magdy]

I am working on a project that aims to evaluate the pronunciation of non-arabic speaker to arabic using sphinx3_aligner and I am stuck into a big problem
I asked someone to pronounce some words with text I wrote for the aligner input and that was the output

SFrm EFrm SegAScr Phone
0 36 -158051 SIL
37 105 -644247 +NONTRANS+
106 109 -96452 l SIL w b
110 112 -125055 w l aa i
113 115 -150550 aa w 7 i
116 118 -146662 7 aa i i
119 122 -46757 i 7 d i
123 126 -58440 d i SIL e
127 146 -90776 +MUSIC+
147 152 -61098 t SIL u b
153 158 -67393 u t f i
159 174 -251284 f u f i
175 178 -79772 f f aa i
179 194 -134562 aa f 7 i
195 206 -33695 7 aa a i
207 223 -194024 a 7 SIL e
224 350 -434997 +NOISE+
351 353 -28280 SIL

in my dictionary which has almost 35 arabic words for now here are the phonemes of the two words he said :

1) w aa 7 i d
2) t u f f aa 7 a

ps: these numberes written in the both words are equivalent to some letters in arabic.

now I know that if the user pronounced the word in a really bad way the aligner won't work. but now It worked and I need to tell the user that he pronounced the word correctly with a grade out of 100, so how could I know if the user pronounced the words perfectly or not. I spent the last few hours reading papers about pronunciation evaluation but I didn't really get what should I do.

Thanks in advance.

[Nickolay V. Shmyrev]

Aligner is not the right tool for pronunciation evaluation. For further info see http://cmusphinx.sourceforge.net/wiki/faq#qhow_to_implement_pronunciation_evaluation

[bassem magdy]

This is the only tool provided by my supervisor to implement the project