The training data corpus was created from ARRL Morse code practice files (text files). In this latest experiment I trained a new Tensorflow based CNN-LSTM-CTC model using 27.8 hours of Morse audio training set (25,000 WAV files - each clip 4 seconds) and achieved character error rate of 1.5% and word accuracy of 97.2% after 2:29:19 training time. This previous blog post covers the new approach of building Morse decoder by training a CNN-LSTM-CTC model using audio that is converted to small image frames. It may not for you though especially if your shack tends to be noisy.I have done some experiments with deep learning models previously. I chose not to as just putting the IPad on top of my radio (see Pic) worked well enough for me. It has been suggested that you connect from your receivers audio output to the audio input on your apple device. Its either the poor cw being sent (remember the local roundtable comment) or you just have to make a very tiny adjustment to the VFO to change the pitch ever so slightly. It worked for me fairly easily as I was used to the tuning aspect from using other decoding softwareĭon’t get discouraged it you don’t get it right the first time.
Read the manual that comes with your radio for the fine tuning on receiving CW signals or just tune by ear for the proper pitch. They send computer generated code and so its about as perfect CW as you can get. Check the ARRL website for times and freqs. I suggest that if you buy the app, get used to it by copying the ARRL code practice on HF. I made 5 more contacts just to prove the 1st one was not a fluke and as a result contesting stations in the following RAC/ARRL sections got the rare “Ontario East (ONE) multi. He acknowledged and then went to calling again. He replied by giving my call and then his exchange 5nn CT
I found a station calling CQ Test with his callsign and tuned the dial on to him… Memory 2 was : 5nn ONE (the contest exchange I had to send)
Just for $h!ts and Giggles I opened up the book of words for my Yaesu FT950 and figured out how to use the built in contest keyer and also how to program the keyers memories. I started reading the screen and even though things were crowded I was able to make out complete contest exchanges. However if you want to use it to pick up the occasional rare DX station or a few extra multipliers in contesting or copy the ARRL Bulletins then this is about the best $15.00 you can spend (in my opinion)… It reads good code, normally that means contest quality computer keyer generated code and its good as what it does.Īs an example last night there was the ARRL 160m cw contest. Remember the old “Garbage in = Garbage out” saying? Well if the person sending the code has a bad fist then…. Right off the bat lets say that if you are looking for a decoder to see what the group on the weekly slow speed CW roundtable net are talking about you will most likely disapointed.
Straight Key or Keyer the choices are yours.
Just to remind you this is a CW Decoding app ONLY (no FT8/PSK or any other sound card modes) and you will need some sort of device for transmitting CW. The cost was about $15.00 Cdn after taxes… It worked right out of the box (or right after installing) on my I Pad. I downloaded the app from the Apple App store and away we went. Searching the app store for Hot Paw Productions should bring you to the app if its still available. When you go to the app store it will tell you if the app will work with the device you are currently using to connect with the app store. This is a app for Apple IPhone and Apple IPad. If anything changes I’ll do a follow up post. This was written 1 day (16 hours) after we bought the app from the Apple App store.