PSK31 vs. CW in white noise

By Lyle Koehler, KØLR

A few years ago I compared the performance of several popular (at the time) LowFER signalling techniques in the presence of "typical" LF noise. It is hard to specify what "typical" really means, because the character of the noise varies greatly from day to night and from location to location. The noise sample used in my previous comparison was recorded off the air, and consisted of a mix of static crashes and electrical noise. Under those conditions, plain old CW offers an advantage over PSK31 for the specific purpose of copying a repeated LowFER beacon identifier in noise. This isn't a fair test of PSK31 vs. CW as modes for real-time communication, but it seemed to make sense because 99 per cent of LowFER operations are one-way beacons sending repeated identifiers.

Many hams using PSK31 have marveled at its ability to maintain copy when the signal was barely audible or even buried in the noise. Does this prove that PSK31 is actually superior to CW as a weak-signal mode? Well, maybe. But PSK31 is typically received using SSB mode, with a relatively wide audio bandwidth, and with an audio pitch that is not necessarily optimum for audible detection of weak signals in noise. Sometimes, the "weak-signal" capability of PSK31 is determined by reducing the receiver's RF or audio gain until the tone is not detectable from the speaker. Sure enough, often the PSK receiving software still produces perfect copy on the screen. However, that is probably not a valid test, since the signal to noise ratio is not affected by reducing RF or audio gain. Even when attenuation is inserted in the receiver antenna line, the signal to noise ratio is not significantly affected until the incoming signal and noise levels are near or below the receiver's noise floor. A better test is to keep the receiver gain at normal levels, have the station at the transmitting end gradually reduce power until the signal is inaudible, and then see what happens to the PSK31 copy.

Although it is difficult to determine exactly how "real" LF noise and propagation will affect various modes, it is very easy to run performance comparisons in the presence of idealized white noise with no propagation-induced amplitude or phase fluctuations. To do that, you go into a good sound editing program (I prefer CoolEdit) and record samples of PSK31 and CW signals. CoolEdit allows you to apply any desired amount of attenuation to the signals, to generate white noise, and to mix the noise and signal. I generated a series of audio test files using this technique. To record the signals, I used the old PSK31SBW software to "transmit" the same text file, first in PSK31 and then in CW. A little bit of fudging was involved, because PSK31SBW sends the CW at a fixed rate of about 18 or 20 WPM -- a little fast for weak-signal work. So I recorded the signals at a CW pitch of 653 Hz with a sample rate of 16,000, and then converted to 11,025 samples/second in CoolEdit. That resulted in a CW pitch of 450 Hz, at a speed of about 13 WPM. I know it sounds like doing things the hard way, but it was quicker than setting up another CW generator program and getting the levels right. Recording of the PSK31 signals was straightforward, using a 1000 Hz pitch with no subsequent conversion. The original PSK31 and CW signals had exactly the same peak-to-peak amplitude, and I then applied 20 dB of attenuation to both signals. A white noise sample with an amplitude somewhat greater than the original signals was generated and then mixed with each of the attenuated signal files.

With these very low signal to noise ratios, PSK31 copy is pretty ragged, and I can barely make out the CW signal in the noise without filtering. But I wouldn't attempt to receive really weak LowFER CW signals without a good filter, so I applied CoolEdit's FFT filter to the CW + noise signal, with a bandwidth of 20 Hz and a center frequency of 450 Hz. It may be difficult to pick out the CW signal in the "ringing" produced by the extremely narrow filter, but it is copiable if you work at it. Does the CW provide better copy than the PSK31? It depends on whether you are better at extracting the Morse characters from the ringing and noise, or picking out word patterns from the garbled text on the screen. Obviously PSK31 is much faster, and if it were slowed down to the same effective throughput as my CW sample, it would do better. Add some error detection and correction, and the performance of PSK could really improve. But Bill de Carle's BPSK modes proved that before anyone had heard about PSK31!

Here are three audio files in MP3 format for those who would like to draw their own conclusions. Each file is three minutes long and a little over 500 kBytes. I would recommend right-clicking on the files and saving them to disk, rather than being at the mercy of whatever player your web browser uses for MP3 files. Remember that the CW signal is at 450 Hz, and the PSK31 center frequency is 1000 Hz. I'd like to hear if anyone achieves one minute of solid copy of the plain-text message using either mode.

PSK buried in noise
Unfiltered CW in noise
Filtered CW in noise

Within a few hours of putting these files on the Web, Dex (W4DEX) submitted the following copy of the filtered CW file:

LF offers true surface wave coverage for several hundred miles, whereas on 160 meters, the lowest frequency ham band in the USA, the true ground wave is usable only to 100 miles or less. Sky wave propagation can extend LF cove...

Stewart, KK7KA, wasn't far behind with perfect "copy", achieved by applying some clever processing techniques to extract a few words of the message, and then searching for that text string on the Internet. He found a summary of an LF presentation I had given to a local radio club, from which the text of the message had been extracted. 100 points to Slytherin for creativity! Stewart also pointed out that the loss of quality in the compressed MP3 file format makes it more difficult to extract the information. In order to provide a fairer test of PSK31, I played back the original .wav files while running the old PSK31SBW software, as well as the latest versions of DigiPan and MULTIPSK. Of the three, MULTIPSK provided the most accurate copy. It seemed to do slightly better when the file was played back in "real time", rather than using the feature of MULTIPSK that allows for high-speed decoding directly from a .wav file. But there wasn't much difference, and even when the file is played back over and over, on some passes the copy looks better than others. It depends on where the errors happen to fall, and how the eye interprets the characters on the screen. Here is the text of the file as received by MULTIPSK:

endred miVszwh#eas onÙMters, the lowest frequency hamt and in the a, the true ground wave is usable only to@00 mil[ or Vss. i kwave propayation\? extend LF cerage O a thousand milen or more, although ionospheric absorptqn tïes a severe toll on LF mulK- oop propagation. The!ropagation path is typically very stable, Tith fade cycles of maey minutes orrven hours, i iciyeasy itfhese low frequencies tpbld equipment with extremely gooh freqmency stability and accuracy. Compt-Z to the H¹or VHF range, LF as pthctically DC, lo construction techniques and componentstite not at all critit al. Unfortunate, with a waveleth on the order of a mile, any reasonable sized antenna will be t#riply inedcient, and the more efficient it is, the narqwer the bandwidraWAtmospheric and man-made jisiincreases as you go lower in frequency, w;h the noise amplitude fŒlowing almone a 1 law.

Note: The text of the CW message is much shorter than the PSK31 message, because of the sending speed and the arbitrary 3-minute duration I chose for the recording. And the beginning of the PSK31 message is cut off, because I started the PSK31 "transmission" before putting CoolEdit into record mode. That seemed like the easiest way to avoid a sound card conflict.

Just for the record, here is the original text:

LF offers true surface-wave coverage for several hundred miles, whereas on 160 meters, the lowest frequency ham band in the US, the true ground wave is usable only to 100 miles or less. Sky-wave propagation can extend LF coverage to a thousand miles or more, although ionospheric absorption takes a severe toll on LF multi-hop propagation. The propagation path is typically very stable, with fade cycles of many minutes or even hours, and it is easy at these low frequencies to build equipment with extremely good frequency stability and accuracy. Compared to the HF or VHF range, LF is practically DC, so construction techniques and components are not at all critical. Unfortunately, with a wavelength on the order of a mile, any reasonable sized antenna will be terribly inefficient, and the more efficient it is, the narrower the bandwidth. Atmospheric and man-made noise increases as you go lower in frequency, with the noise amplitude following almost a 1/f law.