This is a Python3 application Based on PyQt5 for the GUI and PyTorch for the Deep Neural Network. Its purpose is to decode Morse code from the sound coming from an audio device.
At this stage the model (and thus the "program") shows signs of working although has room for improvement. I cannot spend much more time on this since I already dedicated a lot of time to reach this point. However this is Open Source and contributors are welcome to continue the work and bring enhancements. I hope the present materials can serve as a base.
Details on the Neural Network (NN) are given in paragraph H of the Usage section.
This folder contains development Jupyter notebooks. It contains all notebooks from early stage to more elaborated models. The models are trained in the notebooks.
Please check the readme.md file in the notebooks folder for more information.
Some pure python drafts
This is the main application folder containing morseangel.py and its dependencies
You will need Python3 and virtualenv installed in your system. Firstly create and activate a virtual environment:
virtualenv venv
. ./venv/bin/activateInstall prerequisites with pip:
pip install -r requirements.txtStart application:
python ./morseangel.pyThe Neural Network weights are taken from models/default.model you must make sure this file is present.
Just contains the "Exit" item to quit application
The "Device" menu item opens a dialog to choose the Audio input. An audio input must be selected for the program to work.
- 1: Select input device
- 2: Select sample rate among available sample rates for device. As much as possible the 8000 S/s sample rate should be selected or its nearest value.
- 3: Confirm selection and close dialog
- 4: Cancel selection and close dialog
This time line display shows the amplitude of the audio signal
This is the output of the 16k FFT used to find the frequency of the signal peak. The detected peak frequency along with its magnitude in dB is displayed in the legend below the x axis
Use this slider to adjust the Morse code speed in Words Per Minute. You can get help from the envelope signal zoom (F). Yhe optimal length for a dit is 7.69 so the base of a dit pulse should fit in a 10 samples interval. When decodes start to flow the histogram (H) populates and also give an idea of the right setting of WPM. Most amateur radio transmission are done with a WPM around 22~27.
Adjust the value in dB for peak detection.
This is the time line of the detected envelope. Envelope is obtained from the bin of FFT size shown in (I.3) where lies the peak detected by the peak detector (see C). The ±1 bins surrounding the peak bin are also considered (summed up).
The red bars delimit the zoomed view shown in F
The part of envelope between the red bars in (E) is displayed here. A calibrated 13 WPM signal has been used when taking the screenshot so this is the kind of envelope one should be aiming at. Increasing WPM will broaden the peaks.
The decoded text from Morse audio appears here
This is the histogram of element lengths over the length of one character. It is reset at every audio input rate or WPM change. So to reset counts you may just move the WPM slider (D.1) back and forth.
Clearly there are 3 accumulations from lower to higher (left to right):
- Garbage which consists mainly in residual amplitudes of elements not present
- Dits
- Dahs
The decoder is based in splitting lengths into these 3 areas with:
- Lower end for dits: 11 included
- Higher end for dits: 23 excluded
- Lower end for dahs: 25 included
The ideal position of the dits (17) and dahs (32) is displayed with a red and yellow line respectively. The corresponding bin in the histogram is at the right of the line so ideally the peak should appear at the right of the line. In practice having the peak close to the line is good enough. The dit and dah thresholds appear in dashed lines of their respective colors.
One should try to fit the lengths into these limits with the appropriate WPM setting. A calibrated 13 WPM signal has been used when taking the screenshot so this is the kind of histogram one should be aiming at. Increasing WPM will increase element lengths and therefore move peaks to the right.
On strong signals the skill of the operator can also be measured in the shape of the peak. If they are narrow and dahs position is about twice the dits position then timing is correct and regular. As expected this yields better decodes. On weak signals the peaks will broaden inevitably.
This view displays the time lines of the Neural Network output. There are 7 time lines with the corresponding legend:
- Lower line:
- in: input signal
- Middle lines:
- cs: character separator
- ws: word separator
- Top lines:
- e0: first Morse element (dit or dah)
- e1: second Morse element
- e2: third Morse element
- e3: fourth Morse element
- e4: fifth Morse element
Morse characters are decomposed in their constituting elements (the "dits" and the "dahs") that is the "on" periods of the On Off Keying (OOK) signal. The purpose of the NN model is to classify these elements by their relative position in the Morse elements sequence from the start of the character. It has also (of course) to identify the periods of silence into character and word separators. It is not necessary and in fact detrimental to identify the silence between Morse elements. It is limited to 5 Morse elements that is alphanumeric characters plus a few special characters such as +, / and =.
The NN model is based on a LSTM layer. In fact there are two LSTM layers stacked on top of each other (easy to do in PyTorch) and a final Dense (Linear in PyTorch's terms) layer. Thus it takes the imput samples as a stream with a "look back" period corresponding to the longest Morse character possible which is 0 since it is limited to 5 Morse elements. It regurgitates the 7 signals above as sample streams accordingly.
A final purely algorithmic stage does the decoding by identifying character and word breaks using the cs and ws signals and estimating the relative length of the "on" period on each e# element signal. Once the successive "dits" and "dahs" are identified a simple lookup table yields the displayable character that is appended to the decoded text.
Ideally a "dit" period should be represented by 7.69 samples corresponding to the training of the model. For now there is no other way to get close to this value than estimating the Morse code speed in Words Per Minute (WPM) manually. There is an "official" correspondance that states that the period of a "dit" in seconds is 1.2 ÷ WPM.
The preprocessing extracts the envelope based the FFT of the signal with an overlay. This method best preserves the timing of the signal which is essential in Morse coding. Knowing the Morse code speed in WPM the program can compute optimal parameters of FFT length and overlay length to reach 7.69 samples per dit. The FFT size and overlay are displayed in the status line (See next.)
- 1: Audio device selected
- 2: Sample rate in samples per second (S/s)
- 3: Envelope detection FFT size
- 4: Envelope detection FFT overlay
- 5: Device used for Neural Network inference. It can be
cudaif Nvidia GPU can be used elsecpu.
FFT size and overlay is automatically selected for optimal values depending on sample rate and Morse code speed (WPM).