jfb84
commented
6 years ago Hello Theoritically the "score" is the log of the LR, the likelihood ratio between the likelihood of hypothesis "the speech extract corresponds to the model" and the hypothesis "the speech extract was pronounced by someone else". A LR is between ]0 1.0, +infinite. 1.0 is the balanced answer where each hypothesis has the same strenght. LR>1.0 supports more the first hpotheis when LR between 0 and 1 supports more the second one. So a log LR, LLR, is between -infinite to +infinite
Now ALIZE, like all speaker recognition engines, is outputing a raw practical estimation of the LLR. Two steps of "normalization" are usually applied
- at the feature level. In classical ALIZE, the input data are centered and reduced (mean=0 and variance=1 for each coefficient of an input vector)
- at the output level. this process is named "calibration". It is in fact a normalization of the outputted LLR in order to fit better YOUR scenario (speech content, noise, session variability...)
The difference you highlighted comes from the fact that there is no feature normalization applied in this exammple of ALIZE Android. Some the mathematic domain of the inputs are different and the outputted LLr very different (before normalization).
It shows the difference between theory and practical implementation... And shows that someone will have to add normalisation in the ALIZE android example! Best
ra2637
TrungThanhTran
Hi, I'd like to ask how can I evaluate the score in the result? I ran the examples of Alize website and after normalization, the scores are shown as
spk01 -0.269968 0.247707However, the android-alize result shows score as 40 or 110 when I ran it in android app. Can you tell me how to interpret the result?
Thanks!