MagPhase vocoder v2.0 released (April 2018)

[felipeespic]

New in Version 2.0 (April 2018):

• Constant frame-rate support.
• Improved sound quality.
• Two types of post-filter available.
• Selectable number of coefficients for phase features (real and imag).
• Selectable number of coefficients for the magnitude feature (mag).

[m-toman]

Nice.i see you updated the merlin SLT demo "s2" experiment. Perhaps I can free up a couple hours and try it.
Is there anything we should be aware of with the new version?

[felipeespic]

Hi, yes, the "s2" experiment was completely reformatted, now everything is centralised in one main script. Also, you can vary several parameters as wanted.

The function signatures have changed, be aware of that. Other than that, everything should work similarly.

In case you try this version, please, let me know if you have any suggestion or bug you may find.

[m-toman]

I've noticed that it's really clean now, great.
Ran the full demo and the feature extraction is also extremely fast.
Also ran the world demo ("s1") again.

Here are the results, do they sound correct to you?
https://drive.google.com/drive/folders/13BhNi0vCA0ufhTeOfIoWggRZPQZGL4aG?usp=sharing
(the magphase sample is from the magphase ph folder)

[felipeespic]

Thanks! I perceive that it sounds OK, but a bit "wobbly". I think there is something going on with the F0. I am going to check that as soon as I have time. What's your opinion?

[m-toman]

Hmm I think the quality is pretty good but sounds as if there's more power in the lower frequencies, but signal processing is certainly not my strength so I don't think I can help much with that.

I ran the demo as you released it, only changed demo to full. So this seems to be the variable framerate version.

Wonder how it's with 48kHz, but I'd have inject 48kHz recordings first.

[felipeespic]

I realised that the full data demo was using a very simple architecture. So, I just changed it to match what is used in s1. he wobbly effect should be gone now.

Also, did you try with 48kHz data?

[m-toman]

Ah yes, I see and try again.
Interestingly I found that typically the difference between 6x1024 units vs 6x128 is interestingly not that huge, but never tried to reduce the number of layers.

No, unfortunately didn't find the time yet, currently a bit too busy.

[chazo1994]

I have some results of magphase vocoder with lossless and lowdim demo.
https://drive.google.com/file/d/1LONGI3xb4hYV8fMqdS8CDDNCbdFxbtkY/view?usp=drivesdk

The voice quality of lowdim demo still Very bad. I don't know the version which intergrated into merlin is good or bad with my language but it have same algorithm for compress acoustic feature with lowdim demo right?