Here's where I posted the software, wave file, and instructions:

A very cool video Michael. Now I fully understood the basic principles of noise gating. A step further would be to separate scrambled voice and noise from each other thus the noise is gating the voices without being recorded.

Another idea came to my mind. One could fo a similar approach just with the level controlled recording feature in Audacity that actually could do the gating. A downside is the fixed lag-time in Audacity but it would be worth a try.

Yes! We could feed the noise and signal through separate channels like L and R. Then, the signal would be clear as a whistle.

I haven't tried this, yet, but it should doable in my scripts.

One of the caveats I can see is that there may be a little magic (like voices coming through the noise, or signal manipulation) that we would lose.

This is true. The question is if the spirits would agree to adapt. On the other hand multiple voices often interfere even in pure noise, what makes listening to them a challenge. This problem we are facing generally with every technique that gives a continuous stream of voices.

Because the stream of phonemes is fixed (just playing a WAV file) you can compare different runs to see if you're getting different messages. If you're getting the exact same messages each time, then you know the gate isn't set correctly, etc.

I generate the scrambled phonemes with my own Python script, which I've shared here. It ensures that each blip is the same magnitude and clipped so that, in theory, none of them should trip the noise gate without help from an extra noise source.

Getting Python scripts running on your machine requires Anaconda3 and a few module installations, but at least you can see the general idea of what's going on.

phoneme_scrambler2.py

No problem. We have a technique for converting Python scripts to executables. The caveat is each one takes up 700 MB on the user's hard drive. 

Jeff, 

Generally speaking, I haven't spent much time on the phonetic typewriter as I've switched to machine learning assisted "direct voice." However, I agree with you that this work does subtly reveal the abilities and limits of spirit influence for a given hardware system. 

These methods are highly susceptible to audio Pareidolia. 

Have you tried asking for Morse Code ?  ;-)

Eric