about pseudo-clean

[sakiko233]

Hi, thank you for your great work!
I have a question about the generation of pseudo-clean targets in your paper.
I tried to use the pre-trained demucs dns 48 model to denoise some noisy bird vocalizations, but the output was quite distorted and did not sound clean.
Could you please clarify how exactly you generated the pseudo-clean targets for animal vocalizations in your experiments?
Did you use the pre-trained speech enhancement models directly, or did you apply any additional processing or selection to the outputs?
If possible, could you share more details or code snippets about this step?
Thank you very much for your help!

[nkundiushuti]

hi!! I had a PR open with some fixes, one of them is a new domain adaption script. this is useful when the model does not work well on a given dataset.
regarding the pseudo-clean targets, they did not sound great when producing them with the speech enhancement model but they were good enough to train a model. this is also the idea of adapt.py, that through iterative enhancements, you can train a model that works well on your dataset.
train and adaptation have a few options, time scaling and/or noisereducing the targets may produce better pseudo-clean targets.