This article has been published in DUBUS 4/2005 p 19 to 31 and it is protected by copyright. Any reproduction, publishing in the Internet or commercial use only with the written permission from the publisher Verlag Joachim Kraft DUBUS Web page



Speech Processing for SSB Transmitters

By Leif Asbrink, SM5BSZ

http://www.sm5bsz.com/index.htm



The speech processing of modern transceivers is often unsatisfactory in that the out-of-channel interference ("splatter") is much higher than it needs to be. The reason is that the ALC is used for a significant part of the processing (syllabic compression) and that there is no filter after the ALC that removes the out of band signals generated by the ALC. This was discussed in a previous DUBUS article [1].

The reference book on speech processing seems to be Single-Sideband Systems and Circuits by W E Sabin and E O Schoenike of the Collins Radio Company (1987) [2]. It explains the benefits of speech clipping in increasing intelligibility, and that the objective of speech clipping is to increase the average-to-peak ratio of the voice signal. The RF clipper achieves this by removing unnecessary peaks from the voice signal, giving a flat-topped envelope with a greatly improved average-to-peak ratio. But Sabin and Schoenike also warn that some ALC will still be needed after the RF speech clipper. Why is this?

Most DXers know that RF speech clipping process also generates transient intermodulation sidebands (splatter) outside the desired passband, and that these unwanted sidebands then need to be removed before the signal is transmitted. However, it is not so well understood that the well defined flat peak level of the clipped RF signal is a direct result of the presence of these unwanted signals. When the splatter is removed by passing the flat-topped signal through an SSB filter, the signal components that produced the flat top are removed, so new transient peaks must always appear. This phenomenon is named "repeaking". The peak power of an RF-clipped and filtered voice signal is typically 2 dB higher than the peak power of a CW signal that has been clipped to the same level.

If these occasional 2 dB transients were allowed to define the maximum transmitter output (peak envelope power), the average power of the voice signal at all other times would have to be reduced by 2 dB, which is clearly not desirable. That is why Sabin and Schoenike state that a small amount of ALC is needed after the RF clipper and filter, to reduce the transmitter gain by 2-3 dB during the transients while maintaining the full peak envelope power at all other times [2]. They describe how to implement this ALC, and state clearly that the ALC should change the gain by only about 3 dB, with a time constant of about 0.1 second. This is still not a perfect solution (as explained in [1]) because ALC is a form of amplitude limiting that will inevitably reintroduce unfiltered signal energy outside the SSB passband. But if Sabin and Schoenike's recommendation were followed carefully, the ALC-generated splatter would be very small.

This is not done at all in amateur radio equipment [1]. Instead of using only 2-3 dB of ALC to deal with occasional transients, amateur transmitters routinely apply heavy (wide-range) ALC at all times. Also the basics of control loop design - rules like "there should be very little time delay within the complete control loop" - are obviously not generally observed by ham equipment designers. The previous article [1] points out that ALC is also being misused as a wide-range output power control, and to compensate for gain variations in the transmitter due to temperature, frequency and component tolerances. The generated splatter grows rapidly with the ALC range that is being employed. So even things that the Collins company knew about 20-30 years ago are still not being done correctly in today's amateur transmitters.

But today we can use DSP to do even better voice processing than was possible in 1987. This article contains various suggestions that are virtually impossible to achieve in hardware, but could be easily implemented in a modern microprocessor-controlled transceiver with DSP. DSP could also allow us to optimize the processing dynamically, according to the frequency and energy content of the signal at every given moment. For example, one could vary the bandwidth dynamically in front of the clipper, to ensure that the major intermodulation products of the clipped signal will always fall inside the final bandwidth-defining SSB filter. This would eliminate re-peaking and make it unnecessary to use the final 2-3 dB of ALC at all. Better speech processing should not only reduce interference dramatically; it should also help to improve intelligibility a little more than can be achieved by traditional RF clipping.

However, such ideas are as yet unexplored in amateur transceivers. In this article I will explain how AF and RF speech clipping really work, and how the performance of our existing transceivers needs to be improved.

The figures in this article are generated by Linrad. They will clearly show the practical meaning of the mathematical treatment in [2] where descriptions of re-peaking due to filters and Hilbert transforms might be difficult to understand. Linrad also includes a speech processing simulation as part of the setup for the transmit routines. Besides displaying waveforms it allows you to make your own tests to find out how clipping and filtering affects intelligibility with your own voice at different S/N ratios. Linrad can be downloaded from [3].

Clipping of sine waves

The classical argument why RF clipping is much better than audio clipping is illustrated in Figures 1 and 2.


-------------
5. Received
audio 
(distorted).



-------------
4. SSB RF
envelope.




-------------
3. Audio.
Filtered.
then
clipped


-------------
2. Audio.
Filtered.




-------------

1. Original
audio
waveform.


-------------

Figure 1. A sine wave with audio clipping. Note the large distortion in the clipped and received signals, and the non-constant SSB RF envelope (waveform 4).

As you can see, the loudspeaker output signal is quite different between Figures 1 and 2. It is generally assumed that RF clipping is so much better because it retains the correct waveform, a sine wave. In real life this is not true if we discuss intelligibility at the detection threshold. If S/N is degraded to the extent that the tone is barely audible, RF clipping and audio clipping do not sound different. The RF clipper is better for the simple reason that the amplitude is higher at the loudspeaker output, for a given peak envelope power transmitted. With an RF clipper the average power equals the peak power, but with an audio clipper the average power is well below the peak power. For the example in Figures 1 and 2 (a 300 Hz sine wave, 10 dB clipping and a bandwidth of 2.2 kHz), the difference is 3.0 dB in favour of RF clipping. This is a very noticeable difference for weak signals.

But this does not mean that RF clippers are always superior to audio clippers. A human voice is very different from a sine wave, and any conclusion about performance of speech processing drawn from the results of processing a sine wave is highly uncertain and may sometimes be quite incorrect.


-------------
5. Received 
audio.
(undistorted)


-------------
4. SSB RF 
envelope.
Clipped, 
then 
filtered.

-------------
3. Audio. 
Filtered.




-------------
2. Audio. 
Filtered.



-------------
1. Original 
audio 
waveform.



-------------


Figure 2. Sine wave with RF clipping. Note that all audio signals are identical including the received audio and that the SSB RF waveform is constant at the max power level.

Clipping of gaussian pulses

A gaussian pulse is the opposite of a sine wave. It is not periodic at all, and some features of the voice signal are much closer to pulses than to sine waves. Figures 3 and 4 show a gaussian pulse with audio and RF clipping respectively.


-------------
5. Received 
audio.
(not much 
distorted)

-------------
4. SSB RF
envelope.




-------------
3. Audio. 
Filtered,
then 
clipped.


-------------
2. Audio. 
Filtered.




-------------
1. Original 
audio 
waveform.


--------------


Figure 3. A gaussian pulse with audio clipping.





-------------
5. Received
audio
(not much
distorted).


-------------
4. SSB RF
envelope.



-------------
3. Audio.
Filtered.



-------------
2. Audio.
Filtered.




-------------
1. Original
audio 
waveform.


-------------

Figure 4. A gaussian pulse with RF clipping.

Figures 3 and 4 show that audio and RF clipping produce very similar results for a gaussian pulse. Those parts of a human voice signal that are similar to a sequence of pulses should be equally well processed by an audio clipper as by an RF clipper.

Square waves

A square wave is used in [2] to illustrate the re-peaking of a clipped signal. Re-peaking is the reason why audio clippers give less power output compared to RF clippers. Sabin and Schoenike explain this in terms of the Hilbert transform, but a less mathematical explanation can be given as follows. The SSB signal is conveniently generated in a doubly balanced mixer, which is fed by two signals: a constant-amplitude RF signal which is the carrier, and an audio signal. The output from the mixer is the product of the two signals. The carrier is always the same, and therefore the output from the balanced mixer is a RF signal with an amplitude that is proportional to the voice signal. In the special case that the audio signal is a square wave (with a voltage that is +U for 50% of the time and -U for the other 50% of the time) the output from the balanced mixer is a RF signal with constant voltage that is in phase with the carrier when the audio signal is positive and 180 degrees out of phase when the audio signal is negative. In the real world a square wave has rise and fall times that depend on the bandwidth and consequently the 180-degree phase shift is associated with an amplitude that goes through zero for a short time at each phase inversion. Since the output of the balanced mixer is proportional to the audio signal it is clear that clipping the amplitude for the output signal to stay within the limits of a power amplifier can be equally well done on the audio signal as on the RF signal. The reason is that the output from the balanced mixer is a DSB signal. Both sidebands are still present (but no carrier). In order to convert from a DSB to a SSB signal one has to remove one sideband. This means that 50% of the signal energy must be removed. The two sidebands have equal amplitude but their phase differs. It is obvious that the two sidebands can not have the same flat amplitude as their sum, but it is not immediately obvious what they really look like. Figure 5 shows what the RF power looks like for a wideband square wave.

-------------
5. Received
audio
(undistorted).



-------------
4. SSB RF
envelope.



-------------
3. Audio.
Filtered.
then
clipped

-------------
2. Audio.
Filtered.




-------------
1. Original
audio 
waveform.


-------------

Figure 5. Audio clipping of a squarewave produces no distortion of received SSB audio, but it does produce re-peaking of the RF SSB envelope (trace 4).


-------------
4. Received
audio
(distorted).



-------------
3. SSB RF
envelope after 
second filter
(note the
repeaking)

-------------
2. SSB RF 
envelope after
clipping 
(to constant
amplitude)
-------------
1. Original
audio 
waveform.



-------------

Figure 6. RF clipping of a squarewave produces distortion of received SSB audio, and also repeaking of the RF envelope (trace 3).

During the rise and fall of the square wave, the power is very high, but while the square wave voltage is constant, the power is low. In the example of Figure 5 (a 100 Hz square wave in a 20 kHz bandwidth) the peak-to-average power ratio is 9.3 dB. As can be seen from trace 5, the square wave is undistorted in the loudspeaker output of the receiver, but the amplitude is rather low, only 24% of the amplitude of a sine wave at the peak power. Note also that a square wave is not distorted at all by an audio clipper.

Figure 6 shows what happens when the square wave signal of Figure 5 is run through a 20 dB RF SSB clipper. As expected, the peak-to-average power ratio is better, only 4.2 dB, but now the square wave is severely distorted. The RF clipper converts it to something that is halfway between a square wave and a sine wave. In other words, the RF clipper distorts a square wave by replacing the high frequency content of the loudspeaker output with the fundamental and low order components of the square wave. The amplitude of the loudspeaker output is twice as high as without an RF clipper, but the sharp rises and falls of the square wave are reduced by a factor of 2.

This example shows it is false to claim that "audio clipping is much better than RF clipping because audio clipping does not distort the waveform for a square wave". That argument is as false as the similar arguments based on sine waves or gaussian pulses, and for the same reason: the human voice is not any one of those idealized waveforms. Finding out what the optimum processing method is for human speech can only be done with real voice waveforms.

It should be clear by now that the modest loss of peak to average power in Figure 1 is because of the narrow bandwidth. In a bandwidth of 2.2 kHz, the clipped 300 Hz sine wave actually consists of only four components, at 300, 900, 1500 and 2100 Hz. What a clipper does is to change the relation between these components. An audio clipper converts periodic signals towards a square wave in the audio output from the receiver, while a RF clipper converts periodic signals towards a sine wave.

Re-peaking due to the Hilbert transform depends on frequency ratios. The other form of re-peaking, which is due to bandwidth restoration, is also dependent on frequency ratios. The audio or RF envelope waveforms that the clipper produces are flat because all the distortion products are present - these added components are necessary to make the waveforms flat. If a significant amount of the distortion products are then removed because they fall outside the wanted passband, the tops simply cannot be flat any more - there has to be some re-peaking of the final filtered waveform. In [2] it is stated that infinite clipping for a two-tone signal gives a re-peaking that amounts to 2.1 dB. This phenomenon is illustrated in Figure 7.


-------------
5. Received 
audio.
(undistorted, 
same as 1)

-------------
4. SSB RF 
envelope 
after
second 
filter.

-------------
3. Output 
of SSB RF 
clipper.



-------------
2. SSB RF 
envelope of
un-clipped 
signal.

-------------
1. Original 
audio input.



-------------

Figure 7. Two tones, 600 and 1700 Hz with 20 dB RF clipping in a bandwidth of 2.2 kHz.

An RF clipper does not distort the pair of tones shown in Figure 7. The reason is obvious: all distortion products fall outside the passband, so when all distortion products are removed, the shape of the envelope is the same as it was before the clipper. However, the re-peaking amplitude is now 2.08 dB above the clipper level, because the phase of all the intermodulation products was opposite to the signal at the peaks (that is how the peaks came to be flattened, by having those IM products present). At the minimum points of the waveform, the distortion products are in phase with the signal, so they make the minima much narrower.

Figure 8 now shows what happens when the same two tones are clipped with an audio clipper.

The two tones at 600 and 1700 Hz become somewhat distorted when 20 dB of audio clipping is introduced. The reason is that harmonics of 600 Hz fall within the passband. As a consequence the peak-to-average power ratio of the RF signal is degraded by 0.16 dB, although this is a negligible loss of signal power. A comparison between the upper and lower traces in Figure 8 reveals subtile differences. Whether they would lead to an improved or degraded intelligibility if the signal were a part of a voice signal is unclear. Probably it does not matter.


-------------
5. Received 
audio
(slightly 
distorted).

-------------
4. SSB RF
envelope.




-------------
3. Clipped 
audio 
signal,
after 
filter.

-------------
2. Clipped 
audio 
signal.


-------------
1. Original 
audio 
input.


-------------

Figure 8. Two tones, 600 and 1700 Hz with 20 dB audio clipping in a bandwidth of 2.2 kHz.

The fact that audio clipping produces both harmonics and intermodulation products means that re-peaking is larger, because more frequencies fall outside the bandpass filter and are then removed. For this particular signal, re-peaking is 3.75 dB for audio clipping as compared to 2.08 for RF clipping.

There are two obvious ways to reduce re-peaking. One is to use a softer clipper. Like a vacuum tube hi-fi amplifier typically sounds nicer than a transistor amplifier when overloaded, a soft clipper will limit a signal with lower order distortion than a hard clipper. A smaller fraction of the distortion will then be outside the passband so there is less re-peaking. The other way is to use a second clipper with a very small amount of clipping to flatten out the waveform partly. The second clipper would of course have to be followed by another bandpass filter to restore bandwidth again. Procedures like this are of course ridiculously inefficient when implemented in hardware with crystal filters, but in DSP it is easy, and it is possible to gain several tenths of a dB in average power this way for a two-tone test signal at the same peak power.

In a digital system it should be possible to use more direct methods but I am not aware anyone has tried to solve this problem.

Phonetics

To find out what the optimum speech processing is, one has to use real voice waveforms. It is obvious that the very purpose of the speech processor is to distort the waveform. Distortion means dissimilarity between input and output signals. The speech processor may also create differences that are not distortion, such as delaying the signal and changing the amplitude. The distortion is simply the dissimilarity between the transmitter input and the loudspeaker output, expressed as a power ratio (it can also be given in % or decibels). With X(t) for the filtered voice signal containing only the frequency band that we want transferred to the transmitter and Y(t) for the receiver output signal, the distortion D(t) of the speech processor can be written like this:

D(t) = avg[(X(t)*k(t)-Y(t+d))2] / avg[ X(t)2]

where d is the time delay between input and output, and k is the gain. (k may also be allowed to vary slowly with t, at rates 10 times below the lowest frequency within the voice passband or slower.)

The SSB transmission channel is basically linear. Rather than doing the speech processing inside our radio we could pre-distort the voice signal in such a way that it optimizes the usage of the RF channel. With such a pre-distorted signal, no speech processing should be added by the transmitter because the optimum distortion is already introduced. The loudspeaker output at the receiver will then be equal to the audio input signal to the transmitter.

Finding the optimum distortion of a voice signal is a complicated problem. First of all it is pretty obvious that the optimum depends on the situation. When the S/N is very good, distortion may bring unfavourable reports and could cause some loss of intelligibility. In the other extreme when S/N is very low, to the extent that even the slow transfer of e.g. callsigns by use of the phonetic alphabet is unreliable, very severe distortion may actually improve intelligibility. The distortion that optimizes intelligibility at very low S/N ratios may lead to poor intelligibility at high S/N ratios, with sounds that have little similarity to a human voice.

To investigate different forms of speech processing, Linrad can simulate the loudspeaker output from a random sequence of phonetics in white noise. By listening and pressing the corresponding key on the keyboard, one can evaluate the statistics for different kinds of speech processing for different voices and different listeners. As it turns out, speech processing is not critical at all.

Figure 9. Evaluation of Linrad simulation of phonetics for an SSB signal in white noise. 1: No processing, 2 = audio AGC (20dB, no clipping), 3 = audio clipping (20dB), 4 = RF clipping (20 dB), 5 = 2 + 3 + 4 (25dB clipping total), 6 = RF clipping (40dB). The passband is 150-2450 Hz with 6dB/kHz of pre-emphasis.

Figure 9 gives a typical result, the percentage of unknown phonetics read correctly in progressively increasing levels of added noise. As expected, all of the curves show decreasing readability as the noise level increases. The further to the right the whole curve is, the more effective is the speech processing method. Curve 1 in Figure 9 is without any processing at all, and with low enough microphone gain to never saturate the RF channel. All of the speech processing methods (curves 2-6) show a significant benefit over curve 1.

Curve 2 is audio AGC. When I talk into the microphone trying to keep my voice at a constant level, the recorded microphone signal varies by about 10 dB. Automatic gain control can be used to remove this amplitude variation without introduction of any distortion. On the particular test file used for Figure 9, audio AGC improves the readability by about 7.5 dB (curve 2). By using audio clipping rather than audio AGC (curve 3), one introduces some distortion and that gives an additional improvement of about 2 dB. However, it is not possible to get more than about 10 dB improvement by use of straight audio processing. The re-peaking that occurs due to the removal of the undesired sideband limits the average power to about 2 dB below the level that can be reached by use of RF clipping in curve 4 (or audio processing on complex-valued time functions which is equivalent). RF clipping is clearly the best single technique evaluated, although curve 5 shows a small additional improvement through the use of all three processing techniques together, with an overall clipping level of 25 dB. However, curve 6 shows that very heavy RF clipping (40 dB) can have a negative effect on intelligibility at poor S/N ratios (as well as transmitting unacceptable levels of background noise at high S/N).

There is an optimum compression level for clippers. The average output power that falls within the desired passband goes through a maximum when the gain is increased. The phenomenon is illustrated in Figure 10.

Figure 10. Average power (solid lines) and re-peaking (dotted lines) as a function of compression. The audio input is the word "Zulu" and the circles and crosses are for two different male voices.

Compression is the ratio by which the gain for a strong signal is different as compared to the gain for a weak signal. The data in Figure 10 is obtained using RF clipping only. As the gain ahead of the RF clipper is increased causing increased compression, the average output power goes through a maximum (solid curves). The reason is that even when the average power at the output of the clipper stage has almost reached the peak power and cannot increase any more, the peak power losses due to re-peaking (dotted curves) continue to increase with the amount of compression.

The maximum power output is obtained with about 15 dB of compression. With more compression than that, re-peaking increases faster than the average output power at the output of the RF clipper. Transmitters usually use ALC to set the peak power just below the maximum power permitted by the power amplifier, and more re-peaking means more loss of gain. The maximum average power output for the entire phonetics for the word "Zulu" is 2.8 dB below the peak power when all the speech compression is done by RF clipping. Combining AF AGC, AF clipping and RF clipping for a total compression of 25 dB, curve 5 in figure 9 increases the average power for "Zulu" by one more dB to 1.8 dB below the peak power.

That particular word from the phonetic alphabet was chosen for Figure 10 because it is typical of the callsign information that needs to be communicated intelligibly a low S/N. Unfortunately the amount of compression that maximizes the average output power is different for different phonetics. Table 1 gives a few examples.
>br>
------------------------------------------------------------
                                             20dB AF AGC
Phonetics     AF clipper     RF clipper     and RF clipper
              Level Compr   Level Compr     Level Compr
Alfa          -5.50  33     -3.79  33       -3.74  32
Bravo         -6.27  24     -3.95  24       -4.01  25
Charlie       -5.05  33     -4.28  28       -4.27  31
Delta         -6.63  18     -5.04  27       -5.03  30
Yankee        -6.46  28     -5.68  32       -5.37  41
Zulu          -4.36  21     -2.36  14       -2.57  24
-------------------------------------------------------------
Table 1. Largest peak power below PEP obtainable for some different phonetics, and the total compression required to obtain this peak power.

As can be seen in the middle columns of Table 1, the optimum amount of RF clipping spans about 20 dB, if all the phonetics have been normalized in advance by audio AGC to the same peak amplitude. If a fast audio AGC is used to give a fixed compression of 20 dB, this makes a fixed RF compression of about 10 dB close to optimum for all phonetics. "Yankee" is a special case because the sound of the "n" is weak, but with 41 dB of total compression it reaches the peak power while the loss of power for the other syllables is small.

The audible difference between all the different processing methods discussed above is small. The optimum for use in an amateur transmitter is the processing that produces maximum average power with minimum AGC action. Having more AGC than required for maximum syllabic compression will not improve copy, as it will only amplify background noise from cooling fans and similar acoustic noise sources during speech pauses. Besides the fast AGC needed for syllabic compression, there is a need for a slow AGC to compensate for the variation in sound level caused by the operators varying distance to the microphone etc.

The finer details of voice processing as revealed by Figure 10 and Table 1 are not of much practical consequence. This is clearly illustrated by Figure 9. A simple RF clipper with 20 dB compression without any AGC is already very close to optimum if the operator can keep the variations in his voice level within about 10 dB. The few tenths of a dB obtainable by syllabic compression of the audio signal and/or by combining RF clipping with audio clipping will not give a noticeable improvement. AF clipping is not quite as good as RF clipping but the difference is only about 2 dB. The amount of distortion that provides maximum average power depends on the characteristics of the signal. By use of a clipper or very fast AGC that uses information both from the time domain and from the frequency domain, it should be possible to use the optimum compression for each sound, and that might improve the output power by as much as one dB above conventional processing with audio AGC and a RF clipper.

Gain control: AGC and TGC

The introduction to this article pointed out that the correct use for ALC is to compensate for the 2-3 dB of re-peaking caused by the SSB filter following a clipper; but that modern amateur transceivers typically use the same ALC loop to control the power level over a range of 10 dB or even more, and also to compensate for gain variations between bands and due to temperature changes. The ALC function is too critical to be used for all of these purposes at the same time.

When ALC is being used only to compensate for re-peaking, it should have a very fast attack time and about 0.1 second release time. However, it should only have a total gain control range of about 3 dB, with a gain control function that is free from abrupt changes [2]. This system will flatten the waveform each time the level increases suddenly. Such flattening will of course create wideband interference pulses (splatter) but the intensity is low because it does not happen too often. If the ALC is properly designed, the ALC voltage will not decay much between successive peaks within the same syllable and the flattening that occurs in a steady state like in a two-tone test will be negligible.

But if that same ALC system is also used for purposes such as power control, it requires a control range much greater than 3 dB, and successive peaks within the same syllable will be flattened so suddenly that wideband splatter is produced. Unfortunately this seems to be common in amateur radio equipment, and the interference is often made even worse by abrupt changes in the gain control function (as in the FT817). Therefore the large gain variations needed for all other purposes should be done by a different Transmitter Gain Control (TGC) system, which is similar to ALC but with much longer time constants.

Since splatter generated during normal operation with a voice input is not part of the transceiver testing done in amateur publications today, manufacturers do not pay much attention to the transient response of the ALC/TGC system. A modern transceiver could use its microprocessor to set the TGC at a safe initial value, and ramp up the gain to a suitable level in a few milliseconds after the first event when the RF (or audio) clipper has limited the signal level. A microprocessor could also save its previous ALC/TGC settings as a starting-point, so that sudden violent adjustments should rarely be required - instead of happening several times in every syllable, as is presently the case.

Some transceivers use the ALC as the only means for speech processing. An old example is the FT225RD but there are modern examples also. Such transceivers produce wideband splatter even if the gain control function has a good, near logarithmic shape. By introducing clipping on the input side of the SSB filter and disabling the ALC, such transceivers become perfectly clean. If clipping is done on the audio signal (for example by using the clipped audio intended for FM mode) there is a 2 dB loss of average power, although with a more advanced audio processor that processes in Hilbert space, the performance would equivalent with what can be obtained with an RF clipper. Transceivers with particularly badly functioning ALC such as FT817, FT847 and many others can be converted from terrible interference sources to nice-sounding transmitters by a permanent gain reduction to about 2 dB below the power that can be achieved with a sine wave at the microphone input. One way is to inject a suitable (well filtered) DC voltage into the ALC input. This is associated with a very small loss of average output power.

In spite of all this, there is still a possible way to use the ALC as the only means of speech processing. What is required is that the gain control signal is lowpass filtered to not contain any high frequencies. However, to avoid instabilities in the ALC loop one also has to delay the SSB signal by the same amount as the control signal is delayed by the lowpass filter, and the control signal has to be used in a feed forward regulation. Such procedures are easy in DSP but not practical in analog hardware.

A properly designed ALC according to [2] is fine for HF bands where dynamic range requirements are modest. On VHF the splatter that originates in the fast rise time of the ALC voltage is typically the dominating source of interference at separations in the range 15 to 50 kHz. A well designed ALC loop can have some lowpass filtering on the ALC voltage without stability problems. As a result there will be an overshoot like the one in the Kenwood TS2000 in Figure 11. The overshoot is about 0.4 dB and it disappears if the separation between dots and dashes is small. An overshoot like this is absolutely harmless from an interference point of view, although the power amplifier must be capable of handling it with modest compression and without phase modulation or rectification that would lead to a changed operating point. Some amplitude compression is OK, it leads to much less splatter than the sharp corner that a conventional ALC forms. The disadvantage is that the output power is 0.4 dB lower than it could have been with a more intelligent power regulation system.

Figure 11. ALC/TGC with an acceptably small (0.4 dB) overshoot in the TS2000 on 144 MHz. This is the first key-down in CW mode.

Figure 12 is a contrary example of an abrupt AGC. This is the TenTec Orion on 14 MHz. It has a very small overshoot, not so easy to see in the graph, but the carefully shaped waveform is destroyed by ALC action. The discontinuity where the ALC cuts the raised cosine waveform and replaces it by a straight line gives rise to keying clicks. Because the overshoot lasts for a very short time, it creates a wide spectrum as shown in Figure 13. The keying in this example is a continuous sequence of dots at 50 WPM and therefore the ALC action is very small, but normal hand keying looks much worse because the ALC action is stronger after each letter or word space. At about 5 kHz there are strong keying clicks and the excellent phase noise performance of the Orion at close range is hidden under the keying clicks. The TS2000 on 144 MHz (Figures 11 and 14) has much more phase noise, but no keying clicks even at close separations. With a small software update, the Orion could easily become about 40 dB better at a frequency separation of 5 kHz at this high keying rate. At hand keying speeds the possible improvement is about 50 dB.

The Orion curves are shown as one example, to give an idea about the size of the effects. The previous article [1] explains in much greater detail that the strong wideband splatter and keying clicks from many transceivers are the results of design errors - the ALC system does not even function as the designer had intended. Even the results from a wide-range ALC system with immediate attack would be better; but this article has shown there is scope for improvements reaching far beyond that.

Figure 12. The rising edge of a TenTec Orion on 14 MHz when transmitting a continuous string of dots. The carefully shaped waveform is destroyed by ALC action.

Conclusions

Previous articles in this series have shown that, with modern transceivers, the receiver performance has advanced to the point where the largest contribution to inter-station interference is now coming from our transmitters [1]. This article has shown that the speech processing of modern transceivers is often unsatisfactory, because the out-of-channel interference (splatter) is much higher than it needs to be. One of the main reasons is the misuse of ALC. This article has explained how AF and RF clipping really works, and also confirms that only 2-3 dB of ALC is really needed, to control the re-peaking that occurs when clipped signals are passed through the final SSB filter. That information was known to Collins engineers more than 30 years ago [2] but it has been largely ignored by transceiver manufacturers.

References

[1] L Asbrink, SM5BSZ, Real Life Dynamic Range of Modern Amateur Transceivers. DUBUS 2/2005 p. 22-37.
[2] W E Sabin and E O Schoenike, Single-Sideband Systems and Circuits. McGraw-Hill Book Company 1987. ISBN 0-07-054407-7.
[3] http://www.sm5bsz.com/linuxdsp/linroot.htm


Figure 13. The frequency spectrum of the signal shown in Figure 12. Keying clicks are visible as peak hold levels more than 13 dB above the average power spectrum in 2.4 kHz bandwidth.


Figure 14. Spectrum of a hand keyed TS2000 on 144 MHz. This rig produces no keying clicks at all that can be seen in a wideband display like this. That is a natural consequence of the nicely rounded waveform which can be seen in Figure 11.