I'm so confused by the scores in sphinx3 alignment now.
My goal is to detect mistake-reading words or not-well-pronounced words in wave. So i use sphinx3_align. As far as my understanding, if one word's score is low, the word is possibly miss-reading or not-well-reading.
But seems the SegAScr given by function write_wdseg() isn't work as what i am expected. sometime it is positive while sometime it is negative, also it doesn't have high correlation with whether the word is well-pronounced or not.
So i start to track the source-code. in function ms_cont_mgau_frame_eval(), the return-value of function senone_eval() is sometime positive and sometime negative. is senone_eval() calculate the score of current frame belongs one senone?
Also, i don't understand the pre-computation in gauden_dist_precompute(), what does this means
Anybody who can give me some information about this? Are there some documents ?
Thanks a lot
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Well, acoustic score really means nothing serious, it's just the distance to measured by the best hmm. In alignment it's nice measure, but in recognition it's also rescaled and become even less useful. It will not give you an answer on initial task:
My goal is to detect mistake-reading words or not-well-pronounced words in wave. So i use sphinx3_align. As far as my understanding, if one word's score is low, the word is possibly miss-reading or not-well-reading.
To answer on this - you need to introduce "garbage" model and "bad pronounced" words in the dictionary. Then you should build lattice and use confidence score to compare best model with garbage model. Only this comparision will give you and answer if word is correct or not.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm so confused by the scores in sphinx3 alignment now.
My goal is to detect mistake-reading words or not-well-pronounced words in wave. So i use sphinx3_align. As far as my understanding, if one word's score is low, the word is possibly miss-reading or not-well-reading.
But seems the SegAScr given by function write_wdseg() isn't work as what i am expected. sometime it is positive while sometime it is negative, also it doesn't have high correlation with whether the word is well-pronounced or not.
So i start to track the source-code. in function ms_cont_mgau_frame_eval(), the return-value of function senone_eval() is sometime positive and sometime negative. is senone_eval() calculate the score of current frame belongs one senone?
Also, i don't understand the pre-computation in gauden_dist_precompute(), what does this means
Anybody who can give me some information about this? Are there some documents ?
Thanks a lot
I am searching an answer to the same problem. Anybody could give information about the meaning of SegAScr in Sphinx alignment?
Thanks a lot
Well, acoustic score really means nothing serious, it's just the distance to measured by the best hmm. In alignment it's nice measure, but in recognition it's also rescaled and become even less useful. It will not give you an answer on initial task:
My goal is to detect mistake-reading words or not-well-pronounced words in wave. So i use sphinx3_align. As far as my understanding, if one word's score is low, the word is possibly miss-reading or not-well-reading.
To answer on this - you need to introduce "garbage" model and "bad pronounced" words in the dictionary. Then you should build lattice and use confidence score to compare best model with garbage model. Only this comparision will give you and answer if word is correct or not.