Closed vanIvan closed 7 months ago
I have checked my musan/wav.scp, where the keys are all starting from ['noise','speech','music'].
Please check whether the directory structure of the musan dataset (musan.tar.gz) has been changed. You can also check the download link in examples/voxceleb/v2/local/download_data.sh
Okay, I see, thank you for fast checking! It seems that problem is on my side with MUSAN directories layout
Hello!
Based on the following row in
examples/voxceleb/v2/local/prepare_data.sh
:It seems that keys that are generated in
${data}/musan/wav.scp
are wrong: they lack the first subdir that defines type of noise:['noise','speech','music']
:And should be like this:
Due to the missing subdir, in
add_reverb_noise
function when SNR range is picked augmentations SNR fallback to default for all noise files to[0, 15]
.Is it right and expected behaviour? I also have noticed that my models converge to better results when all type of noises are applied with
[0,15]
SNR range.