Re: [wsjt-devel] GFSK for FT8?

[Bill S. <g4...@cl...>]

On 27/04/2019 13:01, Игорь Ч via wsjt-devel wrote:
> Hello Joe,
> .
> Is there any possible losses, e.g. sensitivity or compatibility, at 
> pushing FT8 in software towards GFSK modulation?
> .
> 73,
> Igor UA3DJY

Hi Igor,

the simple answer is yes. But only related to compatibility, for example 
if we elect to use more smoothing of frequency shift discontinuities 
then a decoder designed to be optimal for that will loose sensitivity 
when decoding old style FSK signals. We have some choices if we switch 
to GFSK for FT8, some more practical than others. We can go all out for 
best GFSK performance by matching the decoder to the new GFSK shaped 
signal and accept a loss of sensitivity when decoding old style FSK 
signals, we can try and detect the signal shape and adjust dynamically, 
and so on.

One important area is multi-pass multi-signal decoding with subtraction 
of previously decoded signals, subtraction is most effective if the 
synthesized signal facsimiles to be subtracted are generated in exactly 
the same way that the original transmitted signal was. I.e. it is not 
optimal to generate a GFSK facsimile of a decoded message for the 
purpose of subtraction if the original was actually an FSK signal.

The approach we propose is to restrict the amount of smoothing applied 
to make a GFSK modulated signal to a relatively low amount and to 
perhaps apply an even smaller amount of smoothing to facsimile signals 
used for subtraction, so as to minimize the overall loss of sensitivity 
in a mixed FSK/GFSK environment. Obviously the intent is for all 
stations to move to GFSK modulation eventually as it has clear benefits 
despite the extra complexity of implementation, so any interim measures 
to limit sensitivity loss when decoding FSK modulated signals will be 
biased towards best performance for GFSK decoding. If we take that 
approach I assume the intention will be to move towards decoding purely 
optimized for GFSK modulation over time as users migrate. This in itself 
is not a compatibility issue, rather a quality of implementation issue.

Fortunately the Gaussian function convoluted with a raw FSK signal to 
make GFSK is continuously variable so we can pick one that exactly suits 
our needs and vary it over releases as needed. The final choice of 
Gaussian smoothig functions for transmission and for signal subtraction 
has been made by doing empirical tests using multiple simulated FT8 
signals and using sample data gathered on air. For example measuring how 
many potential decodes are lost decoding real-World FSK modulated 
signals using a decoder optimized for GFSK modulated signals. 
Fortunately for the levels of smoothing we propose this loss of 
sensitivity is relatively small as other effects like noise mixed with 
signals tend to dominate.

For FT4 none of this applies as we are using GFSK modulation from the 
get go and it is simply a case of selecting the amount of smoothing that 
achieves the wanted signal bandwidth characteristics. Note also that 
using GFSK modulation has an impact on transmission when changing a 
message on the fly which is considerably more complex as the whole 
waveform must be pre-calculated whereas with FSK modulation only the 
input symbols need be changed as the synthesized signal is generated 
instantaneously on the fly. There are also implications for ramp up and 
down of the transmission boundaries since the GFSK filtering will 
naturally do this without extra implementation complications.

What is important is that moving to GFSK modulation *will increase 
overall channel capacity* so we hope any temporary loss of sensitivity 
will be outweighed by more potential concurrent decodes per bandwidth 
slot. This should be an attribute that is welcomed by HF users when 
occupancy is often approaching capacity at peak traffic times and at 
other times there will be no loss of performance once GFSK modulation is 
100% adopted.

73
Bill
G4WJS.