Hi Igor, Since this may be of interest to others, I’m copying this reply to the list. First, I’d like to echo what Joe said earlier. If you are running in JT65A mode and if your search range is the same for all cases, then you should see identical results on all platforms and operating systems with the only difference being execution time.
Having said that, I agree that r6052 and r6058 should perform differently on different types of files. In fact, your most recent results are as I would expect. Recall that r6052 uses a standard floating-point correlation function for identifying candidates for further processing. If there are no impairments to the data, then the algorithm used in r6052 should be the most sensitive and the most accurate. As such, it is no surprise that you see better results with r6052 for your “AGC OFF” files and for the -24dB files generated by JTSim. r6058 uses a 1-bit correlation, which was introduced to mitigate the effect of AGC-induced noise “stepdown” (for lack of a better term) at the beginning of the record. It will also mitigate the AGC-induced “stepup” at the end of the record. This approach turns out to be very effective in cases where the AGC effect is severe, as in the example files that you provided. To give an idea of what it does, most of your files produce 100-200 (mostly spurious) candidates with the standard correlation as used in r6052, whereas the 1-bit correlation reduces the number of candidates to 10-30 in most cases. This dramatically reduces execution time, by as much as a factor of 5-10 in some cases. The downside of using 1-bit correlation is a slight loss in sensitivity. Just now, in r6080, I’ve committed my latest attempt at finding the best compromise between robustness and sensitivity. This latest version first uses the full correlation function to identify candidates. If the number of candidates is greater than 50, it then goes back and tries again using the 1-bit correlation. This restores the full sensitivity of the decoder for well-behaved data files (or even for AGC-contaminated data with a narrow search range) while resorting to 1-bit correlation when warranted. If even the 1-bit correlation produces more than 50 candidates (I have not yet seen this happen), then the fallback is to reduce ntrials to 100 to keep the decoder from appearing to “freeze”. For future reference, I just ran r6080 on the -24dB test files using a setting of 8 for random erasure patterns (ntrials=10000), with 2-pass decoding turned off, and obtained 828 decodes. Finally, a comment on your observation that many people have rigs that do not allow the AGC to be turned off. I expect that many such rigs will have an RF gain control. My recommendation is to reduce the RF gain to the point where AGC is activated only on the strongest signals. This will minimize the AGC-induced noise-step effect and will guarantee maximum sensitivity from the decoder. For example, on my TS-480 I usually run with the RF gain set at 80 (out of 100) and only 1-2% of my data files activate the 1-bit correlation option. In any case, this latest version should “just work” whether or not you are able to optimize your AGC settings. 73 Steve k9an > On Nov 11, 2015, at 4:12 PM, Игорь Ч <c_i...@inbox.ru> wrote: > > Hi Steve, > > After long discussion about JT65 decoder in Russian JT65 forum I have been > completely confused by my last test results on comparison of the r6052 and > r6058 releases: > __________________________ > Audio files with TRX AGC OFF : > > release/number of decodes > > 6052/609 > 6058/596 > __________________________ > -24 dB Audio files: > > release/number of decodes > > 6052/765 > 6058/736 > __________________________ > Audio files with TRX AGC ON: > > release/number of decodes > > 6052/863 > 6058/932 > __________________________ > > There is significant difference in number of decoded signals between 6052 and > 6058. > > We have faced a problem: some people do use cheap/classic transceivers where > they may not turn AGC off and they must use JT65 decoder from release 6058. > Other people do use SDRs or more expensive classic transceivers but for some > reasons prefer to keep AGC turned ON - hence they also must use decoder from > r6058. > > But some people have possibility and like to keep AGC turned off and they > would like to get most performance from WSJT-X software, so they do prefer to > use high sensitivity JT65 decoder from r6052. > > What do you think if there is any way to let users to choose appropriate JT65 > decoder, could you discuss this issue with Joe, please? > > 73, > > Igor UA3DJY > ------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
