Author of the paper here.

There is existing work on testing the feasibility of impersonating other person's voice. We discuss them in our related work section at the end of the paper.

I think on the long run, SAS will no longer be a sufficient authentication technique due to advances in speech synthesis. To prolong ZRTP's life we propose usage of sentences instead of words/chars. This is discussed in detail in our best practices section.

This is a fascinating and well elaborated article!

I noticed that UX/UI is important and a guarantee that SAS should increase in length, what are some of the recommendations that you advise to have a good ZRTP implementation ?

Or should we start discussing the fadeoff of ZRTP and a change to something like Matrix protocol or even Signal's one ?

Both Matrix and Signal use WebRTC for VoIP so the content is encrypted by default. Call set up and signaling is also encrypted by default with Signal, and possible with Matrix - it's automatic if the room from which a call is established is already encrypted.

I know Signal attempts to prevent any data leakage by forcing the Opus codec to use a constant bitrate instead of its default VBR -- I'm not sure if Matrix implements anything similar yet.

Instead of using fancy deep learning to fake the voice why not simulate a degraded signal by adding noise and using a good text to speech program (after you have figured out gender ).