The software is

s223 Signal Source Analyzer.

Working on cross-correlation measurement rig (2nd scheme, the 1st depicts my previous efforts):

The idea behind cross-correlation technique is to be able to measure under the noise floor of reference oscillator as well as noise of everything else but the measured oscillator (as long as the noise is stochastic in character thus different in channels).

Since we always have 2nd channel of sound card "for free", it would be shame to ignore it...

So, if we take two FFTs of somewhat similar signals, multiply the FFT's bins at the same frequency offsets - then we reduce bins that are different and leave the bins that are similar between two FFTs. Everything should be done in complex numbers as we'd like to cancel-out the bins with similar magnitude but different phase.

The "basic" hardware part is:

**Few hours later and with LNA:**
The LNA is based on inverting LME49710, with circa 26dB of gain.

Had to change 22uF X7R cap on virtual ground divider as it was too damn microphonic. 1uF MKS did the trick.

Without LNA:

And the result of LNA use: (this pic is with 20dB LNA. I've updated the LNA with extra 6dB):