Signal Processing Needed

[CanadaKim]

I have been following, with great interest the various schematics for receivers, especially ones presented by Andres Ramos. Andres schematics and theories are absolutely flawless, and usually produce amazing results.

The problem seems to be that while we are up to our neck with various receiver schematics, we seem to need to start looking at post signal processing now. We really need to start looking at something that really processes these (usually) croaky, wispy signals into something better, something much more solid and easily understandable.

I do no not specialize in audio theories, as I am a electronics construction-guy, so I am calling out for people that can begin designing the audio processing/noise reduction that is need at this time. 

[Andres Ramos]

Michael Lee is our guru for signal processing using machine learning. His development is far away of anything else we know so far. However even this ist not the ultimate key for everything.

The truth is that we have a big toolbox of means to improve spirit signals depending on the technique used. This starts with filtering processes, going over software noise gating, denoising and ending in esoteric algorithms like Paulstretch. Every tool just improves a particular characteristic  of a signal, not all.

I think this really is the point. Spirit signals can be very inhomogeneous. It's hard to cover all those characteristics with one tool.

[Andres Ramos]

I completely agree with you Jeff. You said exactly the things we observed in our experiments. The kind of "noise" source is predominating the characteristics of the voices. I got clicking voices from clicking sound, rough voices from rough coherer noise, a.s.o.

Maybe this is the reason why the quality of voices gained from white noise isn't good per se even we think it's the ideal entropy source since it contains all frequencies needed with equal levels. But white noise has NO voice characteristics!

One of the best results I ever had was by crumpling mylar foil. This process doesn't produce voices directly but it generates streams of impulses thar are mimicking speech patterns. That was something I discovered as I was deliberately listening to the soft crackling of mylar when you crumple it slowly. The impulses are not equally spaced but come out in bursts like phonems. After applying Paulstretching in Audacity the impulses were turned into intelligible speech.

I think the reason why this works so well is that we need to give them something as close to speech as possible but still not being speech, a speech form without content.