InverseMelScale

audio
1

If i get the below error what should my input tensor look like to the inverse mel scale function:

inv_mel = torchaudio.transforms.InverseMelScale(n_stft=1025)

inv_mel(s)

TypeError Traceback (most recent call last)

in ()
2 plt.imshow(s)
3 inv_mel = torchaudio.transforms.InverseMelScale(n_stft=1025)
----> 4 inv_mel(s)

1 frames

/usr/local/lib/python3.7/dist-packages/torchaudio/transforms.py in forward(self, melspec)
388 “”"
389 # pack batch
→ 390 shape = melspec.size()
391 melspec = melspec.view(-1, shape[-2], shape[-1])
392

TypeError: ‘int’ object is not callable

2

Based on the error message it seems that you are using a np.array instead of a tensor, as the former returns an int from the .size attribute:

x = np.random.randn(10)
x.size()
> TypeError: 'int' object is not callable

x = torch.randn(10)
x.size()
> torch.Size([10])

so you might want to use tensors.

3

Hey, Patrick thanks for getting back. When I use a tensor then i get the following error. I am bit confused what the input should look like given an mel spectrogram of shape: torch.Size([288, 432, 4])

AssertionError Traceback (most recent call last)

in ()
7
8 inv_mel = torchaudio.transforms.InverseMelScale(n_stft=1023)
----> 9 inv_mel(sample)

1 frames

/usr/local/lib/python3.7/dist-packages/torchaudio/transforms.py in forward(self, melspec)
394 freq, _ = self.fb.size() # (freq, n_mels)
395 melspec = melspec.transpose(-1, -2)
→ 396 assert self.n_mels == n_mels
397
398 specgram = torch.rand(melspec.size()[0], time, freq, requires_grad=True,

AssertionError:

4

I guess the input shape might be wrong.
From the docs:

  • n_mels (int, optional) – Number of mel filterbanks. (Default: 128)
  • …
  • melspec (Tensor) – A Mel frequency spectrogram of dimension (…, n_mels, time)

Based on this it seems the n_mels argument is set to 128 by default, while your input tensor has a value of 432 in this dimension.

5

Hey Patrick, so my input tensor can be of shape torch.Size([288, 432, 4]) so what do you think my input tensor shape should be in this case?

Why are there three dots here?: A Mel frequency spectrogram of dimension (…, n_mels, time)
does it imply something?

Thanks

6

The three dots should indicate additional dimensions while the last two dimensions would represent n_mels and time.
Could you set n_mels to 432 in InverseMelScale as it should work:

inv_mel = torchaudio.transforms.InverseMelScale(n_stft=1023, n_mels=432)
x = torch.randn(288, 432, 4)
out = inv_mel(x)
print(out.shape)
7

Hey Patrick,
I actually tried the above step you mentioned before by setting n_mels= 432 and then my cell (jupyter nb) ran for a very long time with no outputs.

8

I think a long runtime is expected, since InverseMelScale uses SGD to solve the mapping.
From the docs:

Solve for a normal STFT from a mel frequency STFT, using a conversion matrix. This uses triangular filter banks.
It minimizes the euclidian norm between the input mel-spectrogram and the product between the estimated spectrogram and the filter banks using SGD.

Corresponding lines of code.

You could change the tolerance to stop the optimization earlier, if needed.

9

Hey Patrick, thanks for getting back. I will try it out again :grinning: