lsnr is not used, what is the use of this variable?

[l2009312042]

lsnr = self.lsnr_fc(emb) * self.lsnr_scale + self.lsnr_offset

i found in the code， this variable is calculated, but not used for loss and other place ,so what is the use of this target？

[Rikorose]

This is a frame-level SNR prediction. This can be used for a voice detection or to turn of the decoder/dfnet for parts where we only have noise (e.g. here)

[l2009312042]

thanks , this make sense .
but i have another question, in your above answer ,when lsnr < -15,the df is on.
but in the paper ,the df is not good for noise part ,those two are contradictory, how to explain this?

[Rikorose]

Oh yea. It should be off, i.e. the other way round. The PR that I linked is not merged and could have some bugs.

But what the paper states is enforced via alpha.
I.e. here

DeepFilterNet/DeepFilterNet/df/deepfilternet.py

Line 279 in ae55008

df_coefs, df_alpha = self.df_dec(emb, c0)

and here

DeepFilterNet/DeepFilterNet/df/loss.py

Line 262 in ae55008

class DfAlphaLoss(nn.Module):

[l2009312042]

thx,i got it
the follow is my understanding of the alpha loss