[200709] How to make a proposal on speaker verification?

[200709] How to make a proposal on speaker verification?

category	item	value
acoustic	feature	40d mel-filter back
frontend	topology	resnet-34 + relu + BN
frontend	loss function	angular prototypical
frontend	pooling	statistic pooling
frontend	embedding dim.	?
backend	scoring	softmax + PLDA?
signal	sampling rate	16 kHz
signal	segment length	2 s
signal	pre-processing	?
data	augmentation	kaldi recipe
data	SAD	No
model	size	1.4 M
model	FLOPs	?
train	batch size	200
train	optimizer	adam
train	learning rate	0.001
train	decaying	5% decay / 10 epochs
train	epochs	500
toolkit	architecture	pytorch