so we break some mining or calibration schemes in an unconstrained environments
facing fifty five combined foundation conditions of various duration and various noise snr types
and reverse specially focusing on low as non short duration conditions
general and calibrations teens employing quality
do you can train conventionally on clean and four data so having three minutes of
speech and no clean speech
and use this on all conditions you facing but you will have a certain calibration
mismatch
calibration loss and what you could otherwise do is you can go for matched calibration
so for each condition your training the colour we just a the calibration parameters but
you will have a lot a high degree of freedom so double the amount of
conditions you need to train as parameters and what we want to focus is
on having a pasta models and no more amount of parameters to train so having
low amount of parameters
so what varying and male and in topic were proposing before was quality measure functions
but use duration does not directly but that's not estimates and low snr are quite
unstable
and we are using audio unified audio characteristics for quality vector estimate
it was already proposed to use be linear kernel combination mattresses for this we of
the reference quality vector and a probe quality vector
but in here we have
that's great amount of parameters of conditions as amount of parameters to train which is
quite high and what we have opposing this function of quality estimates so to use
the cousin of these quality vectors so we got to the degree of freedom of
three witches
karen
arguable
and basically
purpose so and here we depicted calibration mismatch
from having the conventional scheme which is quite high and if we going for the
matched calibration mismatch which is quite low we can approximate quite well as you posted