Reshaping mel-spectrogram and chaning n_mels channels

[Bhanu Teja]

I have a mel-spectrogram with (460,80) shape.

460 length and 80 channels

Now I want to convert that mel into (460,40) how can this be achieved using librosa?

[Vincent Lostanlen]

Hello,

Just set n_mels=40 in the call to melspectrogram

https://librosa.org/doc/0.9.1/generated/librosa.feature.melspectrogram.html

it will be passed to mel

https://librosa.org/doc/0.9.1/generated/librosa.filters.mel.html

I hope it helps

Vincent

--
You received this message because you are subscribed to the Google Groups "librosa" group.
To unsubscribe from this group and stop receiving emails from it, send an email to librosa+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/librosa/9c5f0e0c-99c9-4b15-89fc-93d956a1d184n%40googlegroups.com.

[Bhanu Teja]

Hello,

The problem is I don't have a source wav file. I have a mel with 80 channels and I wanted to convert it to mel with 40 channels by interpolation or something. 

[Vincent Lostanlen]

Hello,

Then check out scipy.signal.interp1d

https://docs.scipy.org/doc/scipy/reference/generated/scipy.interpolate.interp1d.html

Vincent

To view this discussion on the web visit https://groups.google.com/d/msgid/librosa/3fe37c33-d867-45a5-83e6-1da67b4d7410n%40googlegroups.com.

[Bhanu Teja]

Hey  Vincent,

Thanks for looking in this thread. Yes, I tried that as well. My only concern is those 80d values are in mel space so I'm not sure if I can do that. Just wanted to check here if someone tried to do the same here. 

Thanks,

Bhanu

[Vincent Lostanlen]

For a 2x interpolation, i guess this is acceptable.

If you want to be more precise about it, and know all the hyperparameters that came into the design of the 80-band mel-spectrogram, you can give mel_to_stft a try

https://librosa.org/doc/0.9.1/generated/librosa.feature.inverse.mel_to_stft.html

Followed by librosa.feature.melspectrogram with n_mels=40

Sincerely,

Vincent

To view this discussion on the web visit https://groups.google.com/d/msgid/librosa/74bbdecc-2163-452a-a22e-9e0f2dd137adn%40googlegroups.com.

[Brian McFee]

You could also just use a scipy.interpolate.interp1d with mel_frequencies(n_mels=80) as the input positions and mel_frequencies(n_mels=40) as the output positions.  It won't be exactly correct since the filter bandwidths are going to be way off, but depending on your application it might be okay.