eHam

eHam Forums => CW => Topic started by: AG1LE on November 05, 2017, 06:36:02 PM



Title: Latest Machine Learning techniques applied to Morse decoding
Post by: AG1LE on November 05, 2017, 06:36:02 PM
I did some new Morse decoding experiments with Google's TensorFlow open-source software library for Machine Intelligence. I built a LSTM based Dynamic RNN model and trained it with some random text and tested against ARRL 20 WPM materials.  To see the latest results check out:
http://ag1le.blogspot.com/2017/11/tensorflow-revisited-new-lstm-dynamic.html (http://ag1le.blogspot.com/2017/11/tensorflow-revisited-new-lstm-dynamic.html)

PS.  Source code is available in Github https://github.com/ag1le/LSTM_morse (https://github.com/ag1le/LSTM_morse).


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: AE8RS on November 06, 2017, 06:09:16 PM
This looks really interesting. It's been on my list to look into integrating a CW decode engine into a monitoring station. I think this might be a really awesome start.


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: N0PP on November 07, 2017, 06:30:32 AM
Great work. I already work with Tensorflow in another field (image processing). I was planning on looking into CW decoding with deep learning for a while. This will get me started with my own experiments. Thanks for posting!


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: K3UIM on November 19, 2017, 02:18:50 PM
Hopefully my case is a bit different. I became licensed in 1962 and had one great time on 40 cw until 1992 when I received the "me or ham radio" ultimatum.
(Judas Priest but I missed all my homebrew equipment!!!!)

After empty nesting started taking its toll on us I was given the OK and decided to go QRP.

The problem: From running about 20 wpm to falling to about 7 and my bencher wanting to continue doing 25, ... Aarrgghh! I need an electronic means of hearing adjustable cw again to build up my speed. It seems to me that the only cw ops I hear are up in the 20's somewhere and it's frustrating to us "newbies"! Hi Hi.

I don't want to invest in a cassette player, (if they are still to be found), or lp records either, (same reason), so am I just out of luck or what?

Charlie
K3UIM
K3 Usles..... uh ... United In Marriage


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: W1VT on November 19, 2017, 02:32:38 PM
The ARRL has W1AW code practice on in the form of MP3 files if you can't hear the on the air code practice.
G4FON has a morse trainer.
http://www.g4fon.net/


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: VK5EEE on November 19, 2017, 05:27:50 PM
I don't want to invest in a cassette player, (if they are still to be found), or lp records either, (same reason), so am I just out of luck or what?

Charlie
K3UIM
DR OM Charlie, welcome back to CW!!!

No you are not out of luck. All you need is the very browser you are using to read and post in this forum. Go to http://www.lcwo.net and register (takes less than a minute) so you can log in (as it then saves your settings, personal scores etc). There you can use it to send CW to you at any character speed you like, along with any overall speed you like. For example, if you want to listen to 20 WPM (which I think is a good idea so as not to listen to individual dits and dahs too much) but with an overall speed of 10 WPM (thus there will be nice long gaps between letters giving you time to recall the character from your brain memory) then you can do just that.

If you want software to install, indeed as mentioned above, G4FON would be hard to beat, I think anything else is just reinventing the wheel. But for no installation, LCWO.net cannot be beat!


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: K3UIM on November 19, 2017, 07:29:29 PM
The ARRL has W1AW code practice on in the form of MP3 files if you can't hear the on the air code practice.
G4FON has a morse trainer.
http://www.g4fon.net/
I like the W1AW sessions, but they're only 5 minutes each. I've used them and need more time receiving.
What's a "morse Trainer"?
EEE: I'll check out the Icwo thing. Sounds like what I may need!! Thanks.
Charlie



Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: VK5EEE on November 19, 2017, 08:19:55 PM
Do keep us informed, we all want to help anyone coming back to CW. I hope you enjoy the process, it should be ENJOYABLE.


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: W3TTT on November 20, 2017, 10:44:27 AM
Charlie K3UIM
I have a Morse code streaming internet station.  w3ttt.radiostream321.com (http://w3ttt.radiostream321.com)
It is almost always "on".  Please feel free to try it out. 
It is powered by a Java program.  It downloads Google News, filters out all those pesky HTML tags, and other junk, and sends the News text at 12 words per minute. 
I like to get my news via Morse code!

Also remember, Happy Wife Happy Life.  :D

73, Joe W3TTT


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: KM3K on November 20, 2017, 10:55:19 AM
I don't want to invest in a cassette player, (if they are still to be found), or lp records either, (same reason), so am I just out of luck or what?
Charlie
K3UIM
K3 Usles..... uh ... United In Marriage

I'm using the files from K7QO's code-course (http://www.k7qo.com/), which I've put onto a MP3 player that I bought online for $21.
I took a chance on buying it and am well pleased; it is so tiny and very light weight that I could carry it around anywhere.

Over time, I've made this modification just so I could listen to the sounds over and over again to just the letters and the numbers.
I deleted all the intermediary tests because I was interested in taking only the final one that does all twenty-six letters.

I'm now signed up to take the CWOps course next April/May; Jan/Feb is already filled up.

73 Jerry KM3K



Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: KE6EE on November 20, 2017, 04:09:51 PM
I don't want to invest in a cassette player, (if they are still to be found), or lp records either, (same reason), so am I just out of luck or what?

There are inexpensive compact digital recorders which perform very well. A smartphone will do as well, possibly.

LP records and their playing gear are all the rage these days. New LPs are available and old ones play very well on
current gear.


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: K3UIM on November 21, 2017, 07:54:29 AM
Jerry,
Many thanks!! I liked the K7QO item and will be using it as often as possible.
Now I've got to get it from dvd to my ipod!
Charlie


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: W1VT on November 21, 2017, 10:55:59 AM
"Morse Trainer

When I returned to Amateur Radio my CW was decidedly rusty. What I felt I needed was a CW Trainer to get my speed back. I looked around the Web and found an article by Dave Finley N1IRZ on a method of CW training developed by a German psychologist called Ludwig Koch back in the 1930's.

Reading though Dave's article, I decided that it all made a lot of sense. Basically you start off learning the code at the speed you would like to achieve. Unlike the Farnsworth method which seeks to reduce the gap between the letters as you become more efficient, Koch came up with the idea that you should start off just learning two letters at full speed and add an additional letter once you reach 90% proficiency until you have mastered them all. Since you are only learning one new letter or figure at a time, your frustration is significantly reduced."


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: KE6EE on November 21, 2017, 03:30:17 PM
I decided that it all made a lot of sense. Basically you start off learning the code at the speed you would like to achieve...you should start off just learning two letters at full speed and add an additional letter once you reach 90% proficiency until you have mastered them all. Since you are only learning one new letter or figure at a time, your frustration is significantly reduced."

Very sensible and actually a classic learning method. You can sum it up thus: "bite size pieces."

I would add only that if you start learning the code with the rather irrational notion that you would like to achieve 25 or
30 wpm instantly you may experience frustration.

I would select a reasonable speed which has to be slow enough so that you can tell characters apart from one another. If they all sound like gibberish and are indistinguishable, then the speed is too fast for you.


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: VK5EEE on November 21, 2017, 07:13:44 PM
I would select a reasonable speed which has to be slow enough so that you can tell characters apart from one another. If they all sound like gibberish and are indistinguishable, then the speed is too fast for you.
But obviously not so slow that you are counting dits and dahs or making some mental conversion from an image of Morse rather than hearing the overall sound of the letter, i.e. not below 16WPM. That is where a combination of both Koch and Farnsworth, at least for slower speeds may come in handy: i.e. keep the character speed around 18WPM and just reduce the overall speed (i.e. Farnsworth).

Whatever you do, of course, not stressing yourself is KEY ;-)


Title: RE: Latest Machine Learning techniques applied to Morse decoding
Post by: KE6EE on November 22, 2017, 07:14:53 AM
But obviously not so slow that you are counting dits and dahs or making some mental conversion from an image of Morse rather than hearing the overall sound of the letter, i.e. not below 16WPM...
Whatever you do, of course, not stressing yourself is KEY ;-)

Sure thing.

Nothing could be more relaxing while trying to learn something than to constantly trying to calibrate exactly what your "brain" is doing.

Am I committing the felony of counting?

What tragedy may befall me if I drop below 15.9 wpm?

As I've said repeatedly thinking about what you are doing is an ideal way of defeating any attempt at learning.

A prime reason children learn some things so very quickly is that they haven't yet adopted the absurd notion that
there is a right way and a wrong way to do everything.