How to Normalize Melspectrogram?

[Muhammad Faisal]

Hi, can anyone help me regarding how to normalize melspectrogram using librosa.

I convert my audio signal into melspectrogram but i want to normalize it for machine learning.

And is there any way to convert melspectrogram to audio fastly?

[Vincent Lostanlen]

Hello Muhammad,

I recommend Per-Channel Energy Normalization (PCEN)

https://librosa.org/doc/main/generated/librosa.pcen.html


followed by batch normalization


Sincerely,

Vincent.

On 25/10/2021 14:05, Vincent Lostanlen wrote:

Hello Muhammad,

I recommend Per-Channel Energy Normalization (PCEN)

https://librosa.org/doc/main/generated/librosa.pcen.html

followed by batch normalization

Sincerely,

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/be75475a-ded9-4c26-b77a-61d8f5cc1ef7n%40googlegroups.com.

[Muhammad Faisal]

ok basically i am building a text to speech data set where i use melspectrogram as audio feature. but i tried PCEN but the inverse of PCEN Melspectrogram is give very bad audio. is there any way to reverse this normalization after my model predict the Normalized melspectrogram.

[Vincent Lostanlen]

Hello,

Sounds difficult. By definition, any kind of normalization removes some factors of variability in the data so recovering them after the fact is not possible without some side-channel information.

In the case of PCEN, approximating the inverse might actually be doable via convex optimization if you have a good enough initial guess for the denominator ("M").

Another direction is to implement a PCEN-based neural generative model à la WaveGlow.

But both of these entail research questions of their own and are out of the scope of librosa dev.

I hope this helps!

Vincent.

To view this discussion on the web visit https://groups.google.com/d/msgid/librosa/0e499dcb-1912-4427-bb4c-665b36a4ece0n%40googlegroups.com.

[Muhammad Faisal]

So, is there any way I can convert melspectrogram fastly into audio. Librosa Inverse function is too slow. How can i convert it into audio fast.