A Linux/BSD daemon that monitors an audio stream and looks for selcal (Selective Calling; see https://en.wikipedia.org/wiki/SELCAL) calls and emits a timestamp, followed by the selcal code received. The daemon is intended to be as simple and lightweight as possible, and should rely on existing frameworks such as fftw where possible.
SELCAL is a technique that allows a ground radio operator to alert an aircrew that the operator wishes to communicate with that aircraft.
Because of the background noise level experienced on HF radio frequencies, aircrews usually prefer to turn down the audio level of their HF receiver until alerted via SELCAL of a message specifically intended for their aircraft. When the ground station operator wishes to communicate with an aircraft, he enters into the SELCAL encoder the 4-letter code of that aircraft, which is usually included in its flight plan, and transmits that code over the assigned radio channel. All aircraft monitoring that channel receive the SELCAL broadcast, but only those (preferably only one) that have been programmed with that 4-letter code will respond by sounding a chime or otherwise alerting the crew. The crew will then set their volume control higher to listen to the voice traffic and, using ICAO recommended radio procedures, assure that the message is intended for them.
The official specification for the selcal system is found in "ARINC Characteristic 714-6-1990", published on August 15, 1990. The key attributes of selcal codes are as follows:
Selective calling is accomplished by the coder of the ground transmitter sending coded tone pulses to the aircraft receiver and decoder. Each transmitted code is made up of two consecutive tone pulses, with each pulse containing two simultaneously-transmitted tones.
When the ground operator desires to call a particular aircraft, he depresses the buttons corresponding to the code assigned to that aircraft. The coder then keys the transmitter on the air causing to be transmitted two consecutive tone pulses of 1.0 +/- 0.25 sec. duration, separated by an interval of 0.2 +/- 0.1 sec. which makes up the code. Each tone pulse consists of two simultaneously-transmitted tones. The call should consist of one transmitted code without repetition.
The frequency of transmitted codes should be held to +/- 0.15% tolerance to insure proper operation of the airborne decoder.
NOTE: The specification does not indicate the required frequency accuracy of the receiver. Given that [research][3] seems to [show][4] that doppler spreads of 5-20 Hz over polar paths are possible, it seems that as a practical matter, the receiver frequency tolerances have to be more relaxed than the transmitter frequency tolerances. In addition, practical experience has shown that various ground stations do not appear to follow the +/- 0.15% tolerance regardless. An initial estimate of a receiver tolerance of 2-2.5% for the tones should be sufficient to mitigate transmitter, receiver, and sound card frequency errors. This issue more or less disappears if the source of audio is an AM receiver, rather than a SSB one, since the SELCALs are sent as SC-USB, they can be demodulated with an AM receiver with no frequency error, other than the transmitter and ionospheric contributions. Based on a cursory analysis of the live data available, it seems that it is quite easy to have a combination of receiver frequency and soundcard clock errors that sum to 50 Hz or more. This easily puts the tones out of any reasonable detection band.
Overall audio distortion present on the transmitted RF signal should not exceed 15%.
The RF signals transmitted by the ground radio station should contain within 3 dB of equal amounts of the two modulating tones. The combination of tones should result in a modulation envelope having a nominal modulation percentage of 90% and in no case less than 60%.
Tone codes are made up of various combinations of the following tones and are designated by letter as indicated:
Note: The tones are spaced by 10^(0.045)-1 (approximately 10.9%)
| Designation | Nominal Frequency (Hz) | Minimum | Maximum | Width |
|---|---|---|---|---|
| A | 312.60 | 312.13 | 313.07 | 0.94 |
| B | 346.70 | 346.18 | 347.22 | 1.04 |
| C | 384.60 | 384.02 | 385.18 | 1.15 |
| D | 426.60 | 425.96 | 427.24 | 1.28 |
| E | 473.20 | 472.49 | 473.91 | 1.42 |
| F | 524.80 | 524.01 | 525.59 | 1.57 |
| G | 582.10 | 581.23 | 582.97 | 1.75 |
| H | 645.70 | 644.73 | 646.67 | 1.94 |
| J | 716.10 | 715.03 | 717.17 | 2.15 |
| K | 794.30 | 793.11 | 795.49 | 2.38 |
| L | 881.00 | 879.68 | 882.32 | 2.64 |
| M | 977.20 | 975.73 | 978.67 | 2.93 |
| P | 1,083.90 | 1,082.27 | 1,085.53 | 3.25 |
| Q | 1,202.30 | 1,200.50 | 1,204.10 | 3.61 |
| R | 1,333.50 | 1,331.50 | 1,335.50 | 4.00 |
| S | 1,479.10 | 1,476.88 | 1,481.32 | 4.44 |
| Designation | Nominal Frequency | Low | High | Width | Guard |
|---|---|---|---|---|---|
| A | 312.60 | 306.35 | 318.85 | 12.50 | 20.91 |
| B | 346.70 | 339.77 | 353.63 | 13.87 | 23.27 |
| C | 384.60 | 376.91 | 392.29 | 15.38 | 25.78 |
| D | 426.60 | 418.07 | 435.13 | 17.06 | 28.60 |
| E | 473.20 | 463.74 | 482.66 | 18.93 | 31.64 |
| F | 524.80 | 514.30 | 535.30 | 20.99 | 35.16 |
| G | 582.10 | 570.46 | 593.74 | 23.28 | 39.04 |
| H | 645.70 | 632.79 | 658.61 | 25.83 | 43.16 |
| J | 716.10 | 701.78 | 730.42 | 28.64 | 47.99 |
| K | 794.30 | 778.41 | 810.19 | 31.77 | 53.19 |
| L | 881.00 | 863.38 | 898.62 | 35.24 | 59.04 |
| M | 977.20 | 957.66 | 996.74 | 39.09 | 65.48 |
| P | 1,083.90 | 1,062.22 | 1,105.58 | 43.36 | 72.68 |
| Q | 1,202.30 | 1,178.25 | 1,226.35 | 48.09 | 80.48 |
| R | 1,333.50 | 1,306.83 | 1,360.17 | 53.34 | 89.35 |
| S | 1,479.10 | 1,449.52 | 1,508.68 | 59.16 | -- |
fN = 10^((N-1) x 0.045 + 2). For the first tone, N=12, second N=13, etc.
| Designation | Log Frequency | Frequency (Hz) |
|---|---|---|
| A | 2.495 | 312.6 |
| B | 2.540 | 346.7 |
| C | 2.585 | 384.6 |
| D | 2.630 | 426.6 |
| E | 2.675 | 473.2 |
| F | 2.720 | 524.8 |
| G | 2.765 | 582.1 |
| H | 2.810 | 645.7 |
| J | 2.855 | 716.1 |
| K | 2.900 | 794.3 |
| L | 2.945 | 881.0 |
| M | 2.990 | 977.2 |
| P | 3.035 | 1083.9 |
| Q | 3.080 | 1202.3 |
| R | 3.125 | 1333.5 |
| S | 3.170 | 1479.1 |
It is important to note that "real" selcal receivers use AM demodulators to receive the suppressed carrier (SC) transmissions from the ground stations, and so for them, the audio tones received will be effectively the same as those transmitted by the ground stations. The tones are then decoded and used to either signal the user of the call, or to automatically unmute a separate SSB receiver.
For hobbyist uses, there is typically only one receiver, and it is a SSB receiver. This adds additional errors to the received tones, based on offsets in the tuned frequency and the carrier injection (BFO) frequency. One possible solution to this problem is to step back from trying to determine the frequencies of the individual tones and to instead verify that:
- There are two tones present
- The difference in frequency between the two tones matches a pair of known tones in the alphabet
| Tone | Frequency | A | B | C | D | E | F | G | H | J | K | L | M | P | Q | R | S |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 312.6 | 346.7 | 384.6 | 426.6 | 473.2 | 524.8 | 582.1 | 645.7 | 716.1 | 794.3 | 881.0 | 977.2 | 1083.9 | 1202.3 | 1333.5 | 1479.1 | ||
| A | 312.6 | 0.0 | |||||||||||||||
| B | 346.7 | 34.1 | 0.0 | ||||||||||||||
| C | 384.6 | 72.0 | 37.9 | 0.0 | |||||||||||||
| D | 426.6 | 114.0 | 79.9 | 42.0 | 0.0 | ||||||||||||
| E | 473.2 | 160.6 | 126.5 | 88.6 | 46.6 | 0.0 | |||||||||||
| F | 524.8 | 212.2 | 178.1 | 140.2 | 98.2 | 51.6 | 0.0 | ||||||||||
| G | 582.1 | 269.5 | 235.4 | 197.5 | 155.5 | 108.9 | 57.3 | 0.0 | |||||||||
| H | 645.7 | 333.1 | 299.0 | 261.1 | 219.1 | 172.5 | 120.9 | 63.6 | 0.0 | ||||||||
| J | 716.1 | 403.5 | 369.4 | 331.5 | 289.5 | 242.9 | 191.3 | 134.0 | 70.4 | 0.0 | |||||||
| K | 794.3 | 481.7 | 447.6 | 409.7 | 367.7 | 321.1 | 269.5 | 212.2 | 148.6 | 78.2 | 0.0 | ||||||
| L | 881.0 | 568.4 | 534.3 | 496.4 | 454.4 | 407.8 | 356.2 | 298.9 | 235.3 | 164.9 | 86.7 | 0.0 | |||||
| M | 977.2 | 664.6 | 630.5 | 592.6 | 550.6 | 504.0 | 452.4 | 395.1 | 331.5 | 261.1 | 182.9 | 96.2 | 0.0 | ||||
| P | 1083.9 | 771.3 | 737.2 | 699.3 | 657.3 | 610.7 | 559.1 | 501.8 | 438.2 | 367.8 | 289.6 | 202.9 | 106.7 | 0.0 | |||
| Q | 1202.3 | 889.7 | 855.6 | 817.7 | 775.7 | 729.1 | 677.5 | 620.2 | 556.6 | 486.2 | 408.0 | 321.3 | 225.1 | 118.4 | 0.0 | ||
| R | 1333.5 | 1020.9 | 986.8 | 948.9 | 906.9 | 860.3 | 808.7 | 751.4 | 687.8 | 617.4 | 539.2 | 452.5 | 356.3 | 249.6 | 131.2 | 0.0 | |
| S | 1479.1 | 1166.5 | 1132.4 | 1094.5 | 1052.5 | 1005.9 | 954.3 | 897.0 | 833.4 | 763.0 | 684.8 | 598.1 | 501.9 | 395.2 | 276.8 | 145.6 | 0.0 |
Detection of the selcal tones is quite similar to DTMF tone detection, and this has been well documented. There are several approaches available:
- Bandpass filter bank and energy detectors (i.e. analog implementation approach)
- Discrete FFT and energy detection in bins containing selcal tones
- Goertzl algorithm for fast DFT (see https://en.wikipedia.org/wiki/Goertzel_algorithm)
- Chirp-Z transform for DFT (see https://en.wikipedia.org/wiki/Bluestein%27s_FFT_algorithm)
- MUSIC algorithm (see https://en.wikipedia.org/wiki/Multiple_signal_classification)
- ESPRIT algorithm (see https://en.wikipedia.org/wiki/Estimation_of_signal_parameters_via_rotational_invariance_techniques)
- Wavelet transform and convolution (seems highly advanced)
- Q Transform (handles geometric spacing of tones/bins much better)
The implementation should take into account characteristics of the HF radio medium:
- Often poor signal/noise ratio
- Frequent ionospheric and auroral fading and flutter
- Slight doppler due to relative ionospheric motion
These imply that unlike DTMF decoder implementations, a series of measurements should be made during the signal and a final decision determined from statistical analysis of the raw measurements. Based on the highest baseband frequency of approximately 1500 Hz, this means that any audio sampling rate above 4000 samples/second should work fine. Looking at the post-detection stage, we can see that the gap between the two tone groups is specified as 200 mS +/- 100 mS, so in the worst case, the gap could be 100 mS. Following Nyquist and sampling theory, this means that in that 100 mS period, we need to check for the presence or absence of signal at least twice, or once every 50 mS. The actual numbers will vary based on audio sampling rates, but for the typical 44100 samples/second rate of modern sound cards, this leads us to use a block size of 2048 (using a nice round binary number) and a post-processing sample rate of 2048/44100 or 46.4 mS. This also means that we need to window this sample size appropriately and choose an algorithm for tone detection that is "comfortable" with this block size.
After considerable trolling through the Internet, and discarding many approaches (see above) that are oriented toward detecting unknown tones, it seems that theoretically, the ideal approach is to use [cross-correlation][5] of the input signal with known signals to detect their presence, otherwise known as [matched filtering][6].
The figure below shows the cross-correlation between the received SELCAL "JRAE" and each of the 16 tones in the alphabet. The lines are ordered from tone A on the top to tone S on the bottom, and 2.5 to 3.0 seconds of time run from left to right.
Some care must be taken in selecting the block size, as it is a tradeoff between integrating more signal power to improve cross-correlation detection, and added difficulty in detecting the silent period between the tone groups. As mentioned earlier, presumably for direct detection of the shortest possible silent period, blocks should be less than 50 mS, but this would result in requiring detection based on only 15 cycles (312.6 Hz * 0.05 sec) for the lowest frequency tone.
The normal source of sampled audio is the soundcard interface. After some digging, it seems that PortAudio would be a good choice for an audio interface API, since it provides both a degree of platform independence and isolates the application from the various underlying audio frameworks (i.e. ALSA, Pulseaudio, SndKit, etc.). In general, lower sampling rates are preferred due to the reduced processing load, as are fixed point DSP implementations versus floating point implementations.
Clear the result buffer
Clear the tones
Set state to idle
Set the threshold to 50%
While true
Capture one block of audio (~100 mS)
Decimate it by 10
For each tone in the alphabet
Cross correlate the audio with the tone
If the correlation is greater than the threshold
Save the result in the buffer
Else
Add the correlation to the threshold and average
If there are > 800 mS of results in the buffer
If the last result was silence (no tones)
For each tone in the alphabet
If there is a majority of results in the buffer
Then the current tone is detected
If two tones are detected
If the state is idle
Set the state to two tones
Save the two tones
If the state is two tones
Return the four tones
Clear the result buffer
Clear the tones
Set the state to idle