Clifton Laboratories 7236 Clifton Road  Clifton VA 20124 tel: (703) 830 0368 fax: (703) 830 0711

E-mail: Jack.Smith@cliftonlaboratories.com
 

Revision History
15 March 2009: Originally written

I'm not a SSB mode operator, and have yet to make an SSB QSO with either my K2 or K3 transceivers. Accordingly, I didn't have much interest in how Elecraft implemented what it calls, in different places, "speech compression" or "True RF Speech Processor with Adjustable Compression."

However, I'm always interested in how things work, and was curious over the degree to which the K3's transmit intermodulation performance changed with DC supply voltage. That was well over a month ago and I'm still analyzing a mountain of data I collected. Part of the tests, however, involved how the K3's "compression" altered transmitted intermodulation products. This lead me first to investigate how the K3's "compression" functions.

I've intentionally placed quotation marks around the word "compression" because in my personal view this word is misleading in the context in which it's used to describe the K3's speech processing.

Traditionally, two types of processing are employed to increase the "talk power" or perceived loudness of speech:

• Compression—Audio gain is adjusted automatically to bring up low level speech passages and to reduce gain for louder speech passages.
• RF Clipping—Applied at IF and hence translated to the output frequency, the RF envelope is held to a constant value once an (usually adjustable) threshold is reached. this is a form of instantaneous gain limiting.

Compression of the type commonly employed in broadcasting works with relatively slow time constants, with an attack time of perhaps 10 ms and a decay time of several hundred ms. RF Clipping, in contrast, has an extremely short attack and decay time, often on the order of microseconds.

Sabin & Schoenike ed., "Single-Sideband Systems & Circuits" (pp 214-215) explain the difference between clipping and compression:

The two methods of audio processing are clipping and compression. In the clipper the audio signal, after 10 to 20 dB of additional amplification, is simply sliced off at the required level. It is "memoryless" in that the action at any one time does not affect subsequent actions. The strong syllabic peaks, which occur from 5 to 10 times per second, are thus instantly reduced. The weaker parts which lie between the peaks are clipped much less and are therefore relatively stronger. ...

In a compressor the peaks are limited by gain reduction, and the behavior after a peak differs from a clipper. The amplifier gain does not restore immediately but is allowed to recover more slowly in a controlled exponential manner. It has memory. On subsequent peaks, the amount of gain reduction may therefore be much less. the result is that the compressed signal has less distortion but tends to reduce the weaker elements which lie between peaks, especially those which immediately follow the peaks, and therefore tends to be less effective. Compressors are widely used to maintain a constant output level despite variations in voice level.

Of course, if a compressor's time constants are reduced to near zero, the difference between compression and clipping largely disappear. See, e.g., Pappenfus, et al., "Single Sideband Principles and Circuits," Chap. 20, "Signal Processing for SSB Transmission."
 

First, what do we expect to see in an audio compression system? The figure below, from Wikipedia's entry on dynamic range compression, http://en.wikipedia.org/wiki/Audio_level_compression, shows how an audio compressor acts to varying input levels of a steady tone. Above the threshold, an X dB increase in input signal causes a Y dB increase in output signal. The compression ratio is thus X/Y. If there is no compression, then X = Y and the compression ratio is 1:1. If the output increases 1 dB for every 2 dB of input increase, the compression ratio is 2:1, etc.
 

For compression measurements, the 8904A function generator is set to output a 1004 Hz sine wave with an amplitude varying from 1 mV to 3000 mV peak-to-peak under program control. This output is applied to the K3's microphone input through a 30 db attenuator. The K3 is set for 3900 KHz, USB mode, with microphone gain = 30. The output power is measured with an 8482A sensor and 437B power meter. DC power is set to 13.8V, as measured at the K3's DC input connector with the 6652A power supply set for remote voltage sensing.
 

This plot should immediately suggest that normal audio compression is not used by the K3. The input-output curve slope does not appreciably change with input signal for compression settings between 0 and 30, and for compression settings of 35 and 40, extreme slope changes are present.
 

• Input audio can be either generated by the 8904A function generator or a digital recording played from the SX-260 PC.
• The K3's output (RF) can be examined either with a spectrum analyzer or oscilloscope for envelope view.
• The K3's monitor output can be examined on the oscilloscope or saved as a digital recording by  the SX-260.

To see all of the images, click here, or on the waveform image below.

The upper trace (Chan 1 - black) is the input waveform, whilst the lower trace (Chan 2 - blue) is the K3's monitor output. There's 31 milliseconds of processing delay in the K3's DSP stages between microphone input and monitor output, corresponding to a 6 division time shift between the input and output  waveforms. (Click here to see how I measured the time delay.)

At clipping level 25, the rounded high amplitude waveform is converted to nearly a square wave envelope, and the lower amplitude bursts are nearly the same level as the main burst.

The ATCOSIM Air Traffic Control Simulation Speech corpus is a speech database of air traffic control (ATC) operator speech, provided by Graz University of Technology (TUG) and Eurocontrol Experimental Centre (EEC). It consists of ten hours of speech data, which were recorded during ATC real-time simulations using a close-talk headset microphone. The utterances are in English language and pronounced by ten non-native speakers.

I did not use the full 10 hours of data (2.5 gigabytes) but rather collected a limited sub-sample of about 7 minutes of data. Each file is short, running between 5 and 15 seconds normally, and has a limited vocabulary, but rich on numbers and uncommon names. Hence, it's a reasonable analog for amateur radio testing.

The audio sample linked at this page are from two audio fragments, identified as:

• SM3_05_14—Male speaker transcript "speedbird one two nine roger opposite traffic at level three two zero expedite and stop descent at level two nine zero."
• ZF3_02_45—Female speaker transcript "airfrans two six two one identified route trasadingen hochwald morok for further climb call zurich sector one three three decimal four bye bye"

(I have not corrected the supplied transcript for obvious errors, such as "airfrans" for Air France, or  corrected capitalization, such as Zurich for zurich.)

To listen to the source audio, click on the identification in the above paragraphs or click on the buttons below

.

I then ran each of these through the K3, with clipping levels from 0 to 40, in increments of 5, with the results below. Click on the associated button to hear the result or click on the Waveform Image button to see a portion of the speech waveform from the SM3_05_14 sample.

Dave, G3TJP, reminds me that if you hold the shift key down whilst clicking on the button, you can bring up the image in a separate browser window. This allows you to see multiple images side-by-side for comparison.

Clipping Level	SM3_05_14 Sample	ZF3_02_45 Sample	Oscilloscope  Waveform Image
0
5
10
15
20
25
30
35
40

To answer this question, I ran a longer speech sample, consisting of 7 minutes worth of shorter samples from the air-ground corpus, through the K3 at varying levels of compression, looking at the output waveform with the spectrum analyzer on peak hold. The image below shows the composite output of the test at compression levels 0, 10, 20, 30, 35 and 40. In general, increasing the compression level also increases the peak output level, so some of the increase in IMD is related to overall higher output power levels. There is a small increase in transmitted IMD with increased levels of clipping, but the IMD increase is small, amounting to a few dB or less. To see the individual plots, click on the appropriate button below.

This image, as well as the individual compression level images, provides a good idea of the K3's transmitted IMD in the real world of speech, compared with two tone testing.

The image viewed by clicking the buttons below show the output spectrum at the identified clip level in green with the same audio signal with clip level 0 shown in the brown/green trace.

RF clipping, as implemented in the K3, is quite effective at increasing "talk power" without causing major increases in transmitted intermodulation. The resulting audio quality at high levels of clipping may not be suitable for extended rag chew conversations, but will provide useful improvements in contests or other situations where maximum dynamic range of the recovered audio is not the primary consideration.