my little posters about forensic voice comparison
and it's
takes some examples from real casework to describe
a small experiment to find out
what might be a slightly better way of doing things because as i've written here
when
when we do
real world forensic voice comparison
we want to know what the best approaches to use on the circumstances of the
case
and this is really the
main reason for doing this kind of research simply to find out when you've got
of the case the that the case in front of you
what's the best
approach to use
so
it's deals with the situation way you have recurrent probably syllabic words like alright crops
up a lot not too bad crops up a lot hello okay
in both suspect and offender speech samples
and
if you are using semi
semi automatic forensic
speaker recognition methods then a one of the main things that is the model they
the performance
trajectories on the separate syllables like and i
all
but what i wanted to do was to find out whether you get better strength
of evidence
if you don't do that but measure of the formant trajectories over the whole of
the work
a sort of a kind of so to perform well that it from multiple
so that this was tested with some high-level features
from thee formant a pattern and the tonal fundamental frequency in the cantonese would die
which means first
and validated we'd likelihood ratios
so i obviously there's absolutely nothing whatsoever i can
side you about automatic speaker recognition i realise that but they might be some
interesting aspects
concerning the simple a high-level will to the that we that we working and i
do have some interesting things to say that likelihood ratios which you might want to
come check with me about and
the main reason on here goes is to is to how you can help me
out so please come have a chat thank you