Hi Joe and all, I’ve just committed a batch of modified routines associated with my attempts to improve the low-SNR performance of wsjt-x.
I did a number of experiments aimed at identifying why wsjt-x was not performing as well as wsjt on our low-snr test files. I used the jt65 test program and the usual set of 1000 .wav files, each of which contains a single simulated signal with -24dB snr. The original batch of test files was generated using JTSim, and contain 8-bit data at 11.025kS/s. I used the *nix program “sox” to convert the files to 16-bits at 12kS/s. I started by doing some tests that convinced me that the solution would involve decreasing the estimation errors for both f0 and dt. In order to get to the bottom of this with my sanity intact, I ended up making the changes necessary to remove all empirical dt offsets and also all zero-padding at the beginning of data arrays. Negative dt’s are handled by testing indices to ensure that they remain within the bounds of the relevant arrays. At present, the convention is that dt is referenced to the beginning of the wav file. So a transmission that starts at 1.0 second into the file will have a time stamp of 1.0. It will be a simple matter to subtract 1 from the displayed value to make the numbers equivalent to what the current wsjt-x prints out. I gave up tracking the effect of each and every change - but I believe that the most significant ones are: (i) Using full-symbol coherent integration when calculating ccf between data and sync sequence instead of the prior scheme which did half-symbol integration and then summed the powers from the two half-symbols (ii) Added a “peakup” step to the final dt calculated using 1/16-of-a-symbol steps. This gets us to sub-1/16-symbol accuracy which makes a noticeable improvement at low SNR’s. (iii) For reasons that I don’t understand - on my data, it seems to be slightly better to fit for a(1),a(2),a(3), but then ignore a(3) when tweaking the frequency. Since I had to make a significant number of changes, I’ve confined them to renamed routines that have the _exp appended to the normal filename. These modified routines are invoked in the Makefile.jt65, but they are not yet used in 1.6.1. I’d like to know that you can reproduce my improved results and that you are comfortable with the changes before we consider moving them into wsjt-x. Here’s a summary of the improvement on my data: -24 dB files, ntrials=10000 before changes: 639/1000 (this is what the current 1.6.1 gives) after changes: 827/1000 333 hf .wav files recorded on 20m, 2 passes, ntrials=1000 before: 4044 after: 4150 Steve k9an ------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
