By Donn Werrbach, VP of Engineering, Aphex Systems Ltd

Introduction
An argument has been long running as to whether digital audio processors actually generate digital grunge and, for that matter, exactly what is "digital grunge". At Aphex, our listening experience has taught us that analog processing sounds much better than digital, but, other than proposing a number of theories and opinions as to why, we have not undertaken a definitive study on the matter. We have relied upon broadcast engineers and others involved in broadcast sound to discover it by themselves, and to a great extent they have. There is, nevertheless, a "digital imperative" out there that compels many to specify digital everything in a blind-faith belief that, "If it's digital, it's perfect". To our delight, the recent spirited debates between digital audio processing manufacturers has created a stir of claims and disclaims about "digital grunge" that has brought this subject sharply into focus for the industry. While we have no direct stake in this "digital processing grunge war", we do take great interest in it since it summons new attention to the analog versus digital processing question.

Investigation
After a flurry of inquiries from our dealers and customers, we became interested in looking at the output spectrum of our model 2020 and comparing it to the output of an Orban model 8200 digital Optimod. We thought we might find something interesting, and did we ever! I knew the noise floor of the 8200 was noticeably higher than the 2020, but I didn't know exactly what else we were going to discover.

In the course of investigation, I set up a test bench to make measurements on the two processors under identical conditions. Having no predefined test procedures to follow, I created a few that I thought would show various aspects of performance. I was particularly interested in the multiplex output spectrum, and that is entirely the basis for this paper. I did not look at the analog line outputs or the digital audio outputs at all. Perhaps I will do that at a future date and publish the results. I'm sure they will be just as revealing.

I began with a +4dBu analog tone input, paralleled to both processors, which was generated by an Audio Precision System One instrument. The multiplex outputs were matched in level within 0.5 percent using an oscilloscope and a Belar FMS-2 Stereo Monitor and could be switched into a recently calibrated Hewlett-Packard model 3685A Spectrum Analyzer using a straight-wired relay box (no amplifiers). The 8200 analog-input clip level was set for +24dBu and its analog-input reference to +4dBu. The 2020 input reference was set to +4dBu.

I wanted to get measurements that reflect basic performance of the units, so I set them both to their factory "Jazz" presets which provide plenty of leveling and compression but at moderate processing density. On these presets, both processors contain similar available gain, i.e., when all gain reduction is released out, the gain from input to output is about the same. That evens the measurement field for noise floor measurements. The processing intensity, i.e., relative loudness, is also quite similar, preserving the relevance of any comparative measurements of processing artifacts.

The Measurements
I will not individually justify the measurements I took, as I believe they stand on their own merits and are totally relevant and reasonable. Additional tests could have been devised, but I went far enough to see what I needed and stopped. I would certainly invite others out there to make additional tests if they are so inclined. I photographed the screen of the Spectrum Analyzer with each processor and I will forthwith present these photos with a description of the associated test and some comments about what is shown.

Test 1
5KHz Left-Only Input

Aphex 2020	Orban 8200

Model 2020 Spectrum Response
The sweep span is 1KHz to 100KHz. The vertical scale is 10dB/div. The peaks are at 5KHz, 19KHz, 33KHz, 38KHz, and at 43KHz. These are the theoretically correct blips. The 38KHz blip shows incomplete 38KHz rejection but is down better than 70dB which is an outstanding specification for any stereo generator. This is known because the -20dB pilot is peaked at the negative 10dB line. Count down by tens to the line displaying 38KHz and you'll see it is below the negative 70 line. Note that there are NO SPURIOUS BLIPS all the way down to the noise floor.

Model 8200 Spectrum Response
The sweep span and vertical scale is the same as above. Notice the extreme array of spurious frequencies that are present. The tallest spurious blip just hits -47dB (17dB below pilot blip). One favorable to digital audio may be persuaded to argue this is inaudible. However, if all the spurs are added together they create a much more powerful noise composition. Note there are lots of spurs actually below the fundamental frequency of 5KHz. These spurs don't seem to be readily assignable to Nyquist frequencies, and are not harmonically related to the tone, either.

Analysis
I believe this measurement clearly demonstrates what many of us hear as a sizzling or "zippy" sound of the 8200. Remember, 5KHz is in the voice sibilance band, and voices do tend to have a "crunchy" sound through the 8200. I wondered, "If 5KHz would raise up all that hash, what would a higher frequency and lower frequency produce?"

Test 2
100Hz Left-Only Input

Aphex 2020	Orban 8200

Model 2020 Spectrum Response
The sweep span is 100Hz to 100KHz. The vertical scale is 10dB/div. The peaks are at 100Hz, 19KHz, and 38KHz (actually two peaks at 37.9KHz and 38.1KHz merged by the 300Hz resolution bandwidth). These are the theoretically correct blips. The noise floor is free from spurs. The scale is adjusted to place the pilot at -20dB which calibrates the display correctly in dB re:100% modulation. Note the CLEAN spectrum of the model 2020 with no trash or hash.

Model 8200 Spectrum Response
The sweep span and vertical scale is the same as above. An interesting effect is noted. Notice the shape of the noise floor. Actually, the 300Hz resolution bandwidth of the analyzer has blended together thousands of frequency blips to create this curvy floor. Notice that the effect extends clear through the 53KHz stereo spectrum where it drops down and flattens out.

Analysis
This illustrates what happens to pure bass notes in the processors. One would expect low frequencies, hundreds of times below the Nyquist frequency, to be free of junky stuff, but the 8200 still generates it. Granted, the grunge peaks out at a mere -60dB. You could argue it is inaudible, but notice how nice and quiet the 2020 is in comparison. Consider that program material is a mix of many frequencies, all of which will generate digital grunge, and all of that grunge adds together. How difficult is it to imagine it would become audible in a big way?

Test 3
5KHz Mono Input

Aphex 2020	Orban 8200

Model 2020 Spectrum Response
The sweep span is 1KHz to 21KHz. The vertical scale is 10dB/div. The major peaks are at 5KHz, and 19KHz. These are the theoretically correct blips. There are two minor peaks at 10KHz and 15KHz that are distortion harmonics of the 5KHz tone. This distortion amounts to less than 0.1% THD and may be considered quite a low distortion level. The noise floor is free from spurs. The scale is adjusted to place the pilot close to -20dB which calibrates the display reasonably correctly dB re:100% modulation. Note the CLEAN spectrum of the model 2020 with no trash or hash.

Model 8200 Spectrum Response
The sweep span and vertical scale is the same as above. Look at the spurs. There is a large 3rd harmonic of the tone at 15KHz (equal in level to the pilot at -20dB) which no doubt caused by the 8200's clipping. That would amount to 10% THD, approximately. Digital grunge is evident all throughout the spectrum, including frequencies below the fundamental tone. The grunge achieves amplitudes of -55dB.

Analysis
This illustrates that both mono and single-channel tones produce plenty of digital grunge. I want to conclude that this proves the stereo generator is not to blame, but rather the digital audio processing.

Test 4
14KHz Mono Input

Aphex 2020	Orban 8200

Model 2020 Spectrum Response
The sweep span is 1KHz to 21KHz. The vertical scale is 10dB/div. The major peaks are at 14KHz, and 19KHz. These are the theoretically correct blips. The noise floor is free from spurs. The scale is adjusted to place the pilot close to -20dB which calibrates the display reasonably correctly in dB re:100% modulation. Note the CLEAN spectrum of the model 2020 with no trash or hash.

Model 8200 Spectrum Response
The sweep span and vertical scale is the same as above. Look at the spurs. The picture speaks a thousand words.

Analysis
I wanted to see what happens with really high audio frequencies. Here, close to the Nyquist limit, I thought something messy could happen and it surely does. The 2020 is very well behaved, and looks like a perfect signal generator. The 8200 generates a plethora of spurs that veritably "sing their own song". When you listen to a tone sweeping through the higher frequencies, you can clearly hear the beats and birdies represented by these spurs. That is just one more dimension we can add to the definition of audio processing "digital grunge".

Test 5
Noise Floor Check

Aphex 2020	Orban 8200

Model 2020 Spectrum Response
The sweep span is 1KHz to 21KHz. There is a 100hz tone fed to both processors at -20dBu level. This unlocks the gates and lets the gain reduction of the leveling and compression stages release to full gain. The processor is running at maximum net gain, therefore. The 19KHz pilot tone is evident at -20dB. Note the noise floor of the 2020 is incredibly low at -87dB with no glitches or spurs. This noise is purely gaussian analog noise.

Model 8200 Spectrum Response
The sweep span and vertical scale is the same as above. Note the noise floor is only at -60dB, 27dB noisier than the 2020. The noise has a rising frequency characteristic. A quick listen reveals the noise is grainy and harsh, typical of digital quantization hash.

Analysis
This measurement explains why the 8200 can get noisy with low-level program material. The noise is hashy and coarse. In due diligence, I did discover that the 8200's noise floor can be lowered by dropping the analog-input clip level. The noise seems to go down about in proportion to the clip level. However, to get the noise anywhere as low as the 2020, you have to drop the a-i clip level way down where the A/D converter is always clipping the input signals, and that's unacceptable.

My Test Conclusions
The pictures speak for themselves. I believe my investigation clearly shows there is digital grunge in the 8200. The grunge consists of a wide spectrum of unrelated spurious frequencies that occur whenever a signal is present. These spurious frequencies are mostly concealed by the program content, but astute listeners can easily detect their presence generally, even if they cannot identify the specific grunge components.

I have not had an opportunity to test a Cutting Edge Omnia processor or the CRL DP-100 under these same conditions. It would be interesting to compare them  to the 2020 and 8200. Since the DP-100 uses floating-point DSP, while the Omnia and 8200 use fixed-point, I would expect a difference in the noise and spurious levels but I am not sure in whose favor it would go. There is more than the fixed or floating point DSP issue at play: the deployed DSP algorithms are crucial. I hope CRL and Cutting Edge will make their units available for comparative testing.

As for my biggest conclusion out of all of this, you guessed it! Analog is way better, period. There's just no shadow of a doubt about it. Digital control of advanced analog processing, as done in the Aphex model 2020, is the ultimate technology for high audio quality and performance on FM.

One last point: When we introduced the Aphex Digicoder, we exposed a new level of FM signal coverage. Many broadcasters discovered that the ultra clean multiplex spectrum of the Digicoder output actually reduced what had been thought to be multipath interference. I just can't help thinking that a high level of spurious output from a digital audio processor would have to cause more of that multipath distortion.

end.

Trademark notification:
"FM Pro" is a trademark of Aphex Systems Ltd.
"Orban" and "Optimod" are trademarks of Harman International Industries, Inc.
"Belar" is a trademark of Belar Electronics Laboratory, Inc.
"HP" is a trademark of the Hewlett Packard Company.
"Omnia" is a trademark of Cutting Edge