Is there a way to see the probability (according to sphinx) that the word correct. From what I've read the Acoustic Scores (AS) values are probablity densities in log base 1.0001 (s2) or 1.0003 (s3).
E.g. AS of -360000 would then be 1.0001^-360000 = 2.32x10^-16. Or close to nothing... Even if I scale it somehow with the S factor I don't really get any decent probability percentages.
And is there a good way to get rid of words that have very low probability? I tried to change the -inspen but it doesn't seem to make much difference, at least until it's _very_ small.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Anonymous
-
2003-06-10
Magnus -- your second question is one of the classic conundrums of HMM decoding. The raw log probability scores are valid for comparing decoding theories (which is better?), but not for evaluating on an absolute scale (is this a reliable recognition or not?). Evaluating "confidence' in a particular recognition isn't a simple matter.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Is there a way to see the probability (according to sphinx) that the word correct. From what I've read the Acoustic Scores (AS) values are probablity densities in log base 1.0001 (s2) or 1.0003 (s3).
E.g. AS of -360000 would then be 1.0001^-360000 = 2.32x10^-16. Or close to nothing... Even if I scale it somehow with the S factor I don't really get any decent probability percentages.
And is there a good way to get rid of words that have very low probability? I tried to change the -inspen but it doesn't seem to make much difference, at least until it's _very_ small.
Magnus -- your second question is one of the classic conundrums of HMM decoding. The raw log probability scores are valid for comparing decoding theories (which is better?), but not for evaluating on an absolute scale (is this a reliable recognition or not?). Evaluating "confidence' in a particular recognition isn't a simple matter.