I spent some time this morning with the 20 wpm set (20 wpm is about my typing speed!) and was able to do solid copy (essentially zero CER) on the -6 dB and stronger signals. My CER was around 5% at -8 dB, but I fell apart at -10 dB, with only perhaps 50% copy. This is a very preliminary "quick look," based on copying only a minute or so from each file, (I didn't bother with any of the files stronger than + 6 dB).
I will use these data points as reference. Human auditory system seems to have a pretty good filtering capability what comes to Morse code type signals. To achieve similar results with this software I would need to get roughly 16 dB improvement by filtering the original signal that has 2KHz bandwidth. Since the noise in these files is white noise 16 dB translates to roughly 50 Hz bandwidth [16 dB = 10*log10(2000/50)].
I did some work back in January trying to compare the impact of matched filtering vs. FFT filtering in FLDIGI. The plot below shows the big improvement that filtering brings on CER figures:
Once I get this new algorithm integrated better with FLDIGI I can re-run the above CER vs. SNR tests to see how much improvement the Bayesian probabilistic correlator-estimator provides. According to the doctoral thesis I used as the source for this algorithm there should be ~ 4 ... 6 dB improvement.
Thanks again for taking the time for this experiment!