New Bayesian Morse Decoder
In my previous post I shared initial results of the new Bayesian Morse Decoder. I have worked a bit more on this topic to integrate the decoder software to FLDIGI v3.21.75. Dave W1HKJ provided me some space on his web server so I posted full source code including all changes and new source files related to Bayesian Morse decoder in here.
To demonstrate the Bayesian decoding capability I tuned my KX3 on W1AW frequency 3.58105 Mhz at 8:00 PM EST today and let two identical FLDIGI copies of the same v3.21.75 software running from the same audio source. I have a SignaLink USB connected between KX3 and my ThinkPad laptop running Linux Mint OS.
On Fig 1. below I have the legacy SOM decoder on the left side and the new Bayesian decoder on the right . The legacy decoder makes a lot of errors and CW bulletin is not really readable. On the right side I have enabled Bayesian decoder and you can actually read the bulletin, though the lack of proper word spaces make this still a bit difficult to read. In any case Bayesian decoder provides significant accuracy improvement with real life noisy CW signals.
|Figure 1. FLDIGI running legacy and Bayesian Morse decoder
If you choose to compile FLDIGI from the sources yourself please see Figure 2. below. When you go to menu item Configure / Modems / CW / General and click the tick box "Bayes decoding" that will enable the new Bayesian decoder (and disables legacy decoder). You will also notice that CW RX speed indicator (Fig 1. bottom left corner of FLDIGI window) will change more rapidly - this is a feature of the new decoder. It evaluates the instantaneous speed continuously and uses this information as part of the decoding process.
|Figure 2. Enabling Bayesian decoder.
I would be interested in getting your feedback and suggestions how to improve this software. Current version is buggy and it does crash occasionally. Also, I still have not found the root cause of the problem that causes approximately 3% base error rate.
There are number of improvement areas that need work. For example noise estimator ("noise.c") does not seem to be able to keep track of the real noise level. This noise power information is later used by Kalman filters ("kalfil.c") in estimating likelihood of keystate given the observed signals. Bayesian inference is then used in updating posterior conditional probabilities for each new path in ("probp.c"). I am getting intermittent overflows in this section of the code. Any help in this area would be very welcome.
However, given the visible improvement in the decoding accuracy at least during certain conditions (when you have the signal level correct, not too much noise etc.) I am now posting this for alpha testing with hopes that I could get some more eyes to look at the code and find those bugs.