toggle quoted messageShow quoted text
Hm, maybe I can convert the rtlsdr source I have to
handle extio files. That might keep me off the street a few days.
OTOH it would be a more fun challenge to skip the extra dll and
include that code into a sourcexxx.dll. Then you'll be able to
Simonize (ouch) work I kludge into something in your style and
more trouble free.
On 20220102 10:29:21, Simon Brown
see that it would be nice to play with but is not very
practical and requires a lot of effort from me. There’s no
API for the RX888, just raw ADC data and ‘good luck to the
On 20220102 09:44:53, Simon Brown wrote:
A few hours is a few days. Adding serial
number support is straightforward, I have no plans to add
the 64 MHz bandwidth.
Please somewhere in there squeeze in a
few hours to put together the new RX888 knowledge to make it
more useful within SDRC.
On 20220102 08:37:39, Simon Brown
Agreed about the training. I hope to
have the AI included by Easter, then I can dig further.
Still a lot to do to get a nice UI harness working.
Simon Brown, G4ELI
Tensorflow is an excellent system for
this sort of thing. The author of that article used
random words to train the model, but if he used more
traditional ham conversations it would help a lot, at
least for amatuer use including typical abbreviations
and Q signs so they would be better represented and not
read as errors by the model. The other thing that would
be useful would be to send a bunch of known CW text
from an automated keyer and pick up the audio from a
remote SDR, so the model would include fading and other
non-linear impairments that come from real HF paths that
hams deal with all the time. The cherry on top would be
to have hams send a set of known text with keyer, bug or
straight key and capture that over the HF signal path as
If the ML model is trained well, I
would not be at all surprised if it could decode CW
much better than the best human. I have been
surprised as to how amazing the pattern recognition
capabilities of well trained models are. The key is
all in the training.
On Sat, Jan 1, 2022 at 1:14 PM
Simon Brown <simon@...>
And excellent article –
something I want to understand this year.
Simon Brown, G4ELI
There are a huge number of
new tools for using AI today that leverage
modern compute infrastructure. Unfortunately,
many of the public ML tools out there are
optimized for visual image processing, as
opposed to audio or RF SIGINT style uses. But a
number of efforts exist to convert audio or RF
to an image, and then use the image tools to do
the training and generate a model for decoding.
While the training is computationally expensive,
once the model is built, it can be pretty
efficient to run.
There was another one that
used ML, GIMP and RTL-SDR's to do signal
identification using images again, but I can't
seem to find it now.
On Sat, Jan 1, 2022 at 9:55
AM Simon Brown <simon@...>
This actually will be a
step into AI in 2023 – 1988 / 1989 I was
working in AI at SWIFT, sadly the brains
of the operation was on the plane that
crashed at Lockerbie. 30+ years later it’s
time to try and remember the theory and
see how it’s changed.
You will do a great
job. Solving the unsteady CW hand issue
will be interesting. Even some machine
produced CW will be a challenge to
decode. Looking forward to your work on
it in a few months from now. Computers
are amazing but programmers even more
On Jan 1, 2022, at
11:35 AM, Simon Brown <simon@...>
I’ve often said
that a computer can decode better than
the human brain, now I’m trying to
prove it. ‘I may be some time…’
Happy 2022 to
all. May good SDR things happen this
year. Good luck Simon on the CW
decoder. It’s much needed.
On Jan 1, 2022,
at 9:02 AM, Simon Brown <simon@...>
Over here it’s
2022 so a very happy new year to
you all. I’ve updated the current
V3.1 beta and am now spending
January designing a CW Decoder
harness which I hope will help me
develop a good CW Decoder, so for
January and possibly February
you’ll not be hearing much from me
Be good in
- + - + -