search

FrancoisGrondin / mvdrpf

MIT License
3 stars 4 forks source link

Question on the pipeline? #1

Open NormonisPing opened 9 months ago

NormonisPing commented 9 months ago

Dear, Tks. Learned a lot from your amazing Jobs in your profiles.

Regarding pipeline in this project, did I correctly understand?

  1. Data preparation, a. using reverb_target_speech and reverb_inferior as speech and noise spectrum estimation. calculate the mvdr weight. b. dataloader takes three channels of wav: 1. bf_output of target, bf_output of inferior, bf_output of readl target as input, then returns [bf_output_tgt_inf_spec, ideal_mask ... ...]
  2. mvdrpf model Train stage: net in : bf_output_tgt_inf_spec, loss: spec MSE_LOSS with ideal_mask. A bit like single-channel noise reduction.

Two questions:

  1. in practice, how to get the speech and noise spectrum estimation, surely methods in GSC or spatial vad might work.
  2. In MVDR function, steering vector is not assigned, will it affect something?
NormonisPing commented 9 months ago

seems steernet can be used as speech and noise spectrum estimation. mvdrpf oh oh. here, means postfilter. Got it.