joonson / voxceleb_unsupervised Goto Github PK

Hi, thank you for your amazing work. I'm wondering whether there is an instruction for loading both the image face frames as well as the speech segments.

Hello, thank you for good resource.
I guess, in the paper, reversing gradients is only done when the embedding training phase.
However, in the source code, it seems that the discriminator training phase also goes through the GRL layer, so I wonder why.

Thanks for the perfect job. Can you describe how to get 1000 pre-computed RIR filters specifically?

Thanks for your work. When I reproduce your paper, i can't achieve the result in your paper: AP+Nosie+RIR eer=9.56%,
AP+AAT+Noise+RIR=8.65%, I only get the result: AP+Nosie+RIR eer=11.78%, AP+AAT+Noise+RIR=10.08%. Is there anything else I need to pay attention to during training?

def gen_echo(ref, rir, filterGain):

rir     = numpy.multiply(rir, pow(10, 0.1 * filterGain))    
echo    = signal.convolve(ref, rir, mode='full')[:len(ref)]   

return echo

in this function, rir data type is float32, but ref data type is int16.

can convolve the different data type data?

Is there any problem in terms of signal processing?

Hello,

for ii in range(0,it-1):
    if ii % args.test_interval == 0:
        clr = s.updateLearningRate(args.lr_decay) 

It seems to be one more learning rate decay when ii is zero.
Thank you.

What's the best unsupervised method do you know? And does anyone reproduce the AAT result?

Hi. Thank you very much for the good material!

In your code, I can't find the augmentation part including both the RIR and noise employed in your paper.
In addition, I'm wondering if the resource of the Augmentation Adversarial Training will be released or not.

Thank you.