I can think of its usefulness in case of speaker verification, where we have a threshold and by score normalisation we are moving a true and imposter scores as far as possible from each other, which leads to bigger margin for choosing a fair threshold.
But in speaker identification there is another scenario, where we don't need any threshold and we just do care about the model which archives the highest likelihood score.!
All the the best
Saeid
What if, the person who is speaking is not in your list? So, you need a minimum threshold to discard some possible errors. After that, you need to compare scores and find the best match.
Score normalization is related to score calibration. Well calibrated scores can be interpreted as loglikelihoods, and therefore used for more complex analisys of them (as, for example, detecting impostors, given aplication parameters such as priors and error costs)
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers