Hi Joe, > Correct False > Program Decodes Decodes Decoder > ------------------------------------------ > JT65-HF 2329 24 BM + kvasd > WSJT-X r5912 2249 0 BM + kvasd > WSJT-X r5955 2114 0 BM + sfrsd > WSJT-X r5955 1816 0 BM only > > JT65-HF yielded 104 more decodes that r5912, but 24 of them were false. None > of the runs with WSJT-X produced any false decodes. It's clear that in the > necessary trade-off between maximum number of good decodes and minimum number > of false decodes, these versions of WSJT-X have been tuned less aggressively > than JT65-HF.
I did an experiment similar to yours using a batch of 333 .wav files recorded on 20m, mostly under daytime with fairly crowded band conditions. Program Good Bad Soft Decoder -------------------------------------------------------- 1. WSJT-X r5922 3125 0 574 BM+kvasd (18.4% kvasd) 2. WSJT-X r5922 3123 2 572 BM+sfrsd (18.4% sfrsd, ntrials=10000) 3. WSJT-X r5922 2551 0 0 BM only 4. WSJT-X r5955 2704 0 482 BM+sfrsd2 (17.8% sfrsd, ntrials=5000) 5. WSJT-X r5955 2222 0 BM only The column labeled “Soft” is the number of good decodes that were obtained by soft-symbol decoding. For cases 1-3 I set the ncount threshold to 0, so these are really r5922-dirty. For case 5, I disabled the sfrsd2 decoder by setting the erasure probability “gain” to 0.0 (the original value was 1.3). Conclusions - - sfrsd and kvasd are equally good in cases 1 and 2, but sfrsd may be producing more bad decodes. - fraction of total decodes produced by sfrsd is 0.182 in case 2 and 0.178 in case 4 - so sfrsd2 is performing roughly as well as it did in r5922 when it was a stand-alone program. - r5955/r5922 ratio of BM-only decodes is 0.87, suggesting that our latest version is sending only 87% as many vectors to the decoder. Steve k9an ------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
