speechbrain: interfaces.py separate_file() not working properly?

Specs: Windows Version 10.0.19042 Build 19042, Python 3.8.3

I’m following the Source Separation tutorial that can be accessed from this page.

Since the audio I have is already mixed, I tried to use model.separate_file(), based off the HuggingFace speechbrain/sepformer-wsj02mix code. There are two issues with that:

1. CPU, RAM and disk usage on my PC shoot up into the high nineties, causing the computer to freeze.
2. After the computer unfreezes (presumably, after the calculations are finished), an error is thrown:

Traceback (most recent call last):
  ...
    est_sources = model.separate_file(path='data/test/audio/speech.wav')
  File "C:\...\venv\lib\site-packages\speechbrain\pretrained\interfaces.py", line 710, in separate_file
    est_sources = self.separate_batch(batch)
  File "C:\...\venv\lib\site-packages\speechbrain\pretrained\interfaces.py", line 669, in separate_batch
    est_mask = self.modules.masknet(mix_w)
  File "C:\...\venv\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\...\venv\lib\site-packages\speechbrain\lobes\models\dual_path.py", line 1124, in forward
    x = self.dual_mdl[i](x)
  File "C:\...\venv\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\...\venv\lib\site-packages\speechbrain\lobes\models\dual_path.py", line 975, in forward
    inter = self.inter_mdl(inter)
  File "C:\...\venv\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\...\venv\lib\site-packages\speechbrain\lobes\models\dual_path.py", line 597, in forward
    return self.mdl(x + pos_enc)[0]
RuntimeError: The size of tensor a (2830) must match the size of tensor b (2500) at non-singleton dimension 1

My guess is that the data that separate_file() passes to separate_batch() is incorrect:

source, fl = split_path(path)
path = fetch(fl, source=source, savedir=savedir)
batch, _ = torchaudio.load(path)
est_sources = self.separate_batch(batch)

Does resampling the data have something to do with it?

If, instead of using separate_file(), I write:

mix, fs = torchaudio.load('data/test/audio/speech.wav')
resampler = torchaudio.transforms.Resample(fs, 8000)
mix = resampler(mix)
est_sources = model.separate_batch(mix)

as suggested in the aforementioned tutorial, the computer doesn’t freeze and est_sources is a torch.Tensor with torch.Size([1, 471272, 2]), which looks like expected behavior to me.