May 11, 2012, by Ethan Winer
Jonathan: “You lied first.”
Jack: “No, you lied to me first.”
Jonathan: “Yes, I lied to you first, but you had no knowledge I was lying. So as far as
you knew, you lied to me first.” — Bounty hunter Jack Walsh (Robert De Niro) arguing with white-collar criminal Jonathan Mardukas (Charles Grodin) in the movie Midnight Run
When it comes to audio fidelity, the four standard parameter categories can assess any type of audio gear.
Although published product specs could tell us everything needed to evaluate a device’s transparency, many specs are incomplete, misleading, and sometimes even fraudulent.
This doesn’t mean specs cannot tell us everything needed to determine transparency—we just need all of the data.
However, getting complete specs from audio manufacturers is another matter. Often you’ll see the frequency response given but without a plus/minus dB range. Or a power amp spec will state harmonic distortion at 1 kHz, but not at higher or lower frequencies where the distortion might be much worse. Or an amplifier’s maximum output power is given, but its distortion was spec’d at a much lower level such as 1 watt.
Lately I’ve seen a dumbing down of published gear reviews, even by contributors in pro audio magazines, who, in my opinion, have a responsibility to their readers to aim higher than they often do. For example, it’s common for a review to mention a loudspeaker’s woofer size but not state its low-frequency response, which is, of course, what really matters.
Audio magazine reviews often include impressive-looking graphs that imply science but are lacking when you know what the graphs actually mean. Much irrelevant data is presented, while important specs are omitted. For example, the phase response of a loudspeaker might be shown but not its distortion or off-axis frequency response, which are far more important.
I recall a hi-fi magazine review of a very expensive tube preamplifier so poorly designed that it verged on self-oscillation (a high-pitched squealing sound). The reviewer even acknowledged the defect, which was clearly visible in the accompanying frequency response graph.
Yet he summarized by saying, “Impressive, and very highly recommended.” The misguided loyalty of some audio magazines is a huge problem in my opinion.
Even when important data are included, they are sometimes graphed at low resolution to hide the true performance. For example, a common technique when displaying frequency response graphs is to apply smoothing, also called averaging. Smoothing reduces the frequency resolution of a graph, and it’s justified in some situations. But for loudspeakers you really do want to know the full extent of the peaks and nulls.
Another trick is to format a graph using large, vertical divisions. So a frequency response line may look reasonably straight, implying a uniform response, yet a closer examination shows that each vertical division represents a substantial dB deviation. The graphs in Figures 1—3 below were all derived from the same data but are presented with different display settings.
For this test I measured the response of a single loudspeaker in a fairly large room with a precision microphone about a foot away. Which version looks more like what loudspeaker makers publish?
Figure 1: Loudspeaker response as measured, with no smoothing.
Figure 2: The exact same data but with third-octave smoothing applied.
Figure 3: The same smoothed data as in Figure 2, but at 20 dB per vertical division instead of 5 dB, making the loudspeaker’s response appear even flatter.
“Empirical evidence trumps theory every time.”
Noise measurements are fairly simple to perform using a sensitive voltmeter, though the voltmeter must have a flat frequency response over the entire audible range.
Many budget models are not accurate above 5 or 10 kHz.
To measure its inherent noise, an amplifier or other device is powered on but with no input signal present; then the residual voltage is measured at its output.
Usually a resistor or short circuit is connected to the device’s input to more closely resemble a typical audio source.
Otherwise, additional hiss or hum might get into the input and be amplified, unfairly biasing the result.
Most power amplifiers include a volume control, so you also need to know where that was set when the noise was measured. For example, if the volume control is typically halfway up when the amplifier is used but was turned way down during the noise test, that could make the amplifier seem quieter than it really is.
Although it’s simple to measure the amount of noise added by an audio device, what’s measured doesn’t necessarily correlate to its audibility. Our ears are less sensitive to very low and very high frequencies when compared to the midrange, and we’re especially sensitive to frequencies in the treble range around 2 to 3 kHz.
To compensate for this, many audio measurements employ a concept known as weighting. This intentionally reduces the contribution of frequencies where our ears are less sensitive. The most common curve is A-weighting, as shown in Figure 4.
Figure 4: A-weighting intentionally reduces the contribution of low and very high frequencies, so noise measurements will correspond more closely to their audibility.
In the old days before computers were common and affordable, harmonic distortion was measured with a dedicated analyzer. A distortion analyzer sends a high-quality sine wave, containing only the single desired frequency with minimal harmonics and noise, through the device being tested.
Then a notch filter is inserted between the device’s output and a voltmeter. Notch filters are designed to remove a very narrow band of frequencies, so what’s left are the distortion and noise generated by the device being tested. Figure 5 shows the basic method, and an old-school Hewlett-Packard distortion analyzer is shown in Figure 6.
Figure 5: To measure a device’s harmonic distortion, a pure sine wave is sent through the device at a typical volume level. Then a notch filter removes that frequency. Anything that remains are the distortion and noise of the device being tested.
Figure 6: The Hewlett-Packard Model 334A Distortion Analyzer. (Photo courtesy of Joe Bucher.)
Intermodulation distortion is measured using two test tones instead of only one, and there are two standard methods. One method sends 60 Hz and 7 kHz tones through the device being tested, with the 60 Hz sine wave being four times louder than the 7 kHz sine wave.
The analyzer then measures the level of the 7,060 Hz and 6,940 Hz sum and difference frequencies that were added by the device. Another method uses 19 kHz and 20 kHz at equal volume levels, measuring the amplitude of the 1 kHz difference tone that’s generated.
Modern audio analyzers like the Audio Precision APx525 shown in Figure 7 are very sophisticated and can measure more than just frequency response, noise, and distortion. They are also immune to human hearing foibles such as masking (1), and they can measure noise, distortion, and other artifacts reliably down to extremely low levels, far softer than anyone could possibly hear.
Figure 7: The Audio Precision Model APx525 Audio Analyzer. (Photo courtesy of Audio Precision)
Professional audio analyzers are very expensive, but it’s possible to do many useful tests using only a Windows or Mac computer with a decent-quality sound card and suitable software. I use the FFT feature in Sony’s Sound Forge audio editing program to analyze frequency response, noise, and distortion.
For example, when I wanted to measure the distortion of an inexpensive sound card, I created a pure 1 kHz sine wave test signal in Sound Forge. I sent the tone out of the computer through a high-quality sound card having known low distortion, then back into the budget sound card, which recorded the 1 kHz tone. The result is shown in Figure 8. (Other test methods you can do yourself with a computer and sound card are described in Chapter 22.)
Figure 8: This FFT screen shows the distortion and noise added by a consumer-grade sound card when recording a 1 kHz sine wave.
As you can see in Figure 8, a small amount of high-frequency distortion and noise above 2 kHz was added by the sound card’s input stage. But the added artifacts are all more than 100 dB softer than the sine wave and so are very unlikely to be audible.
Low distortion at 1 kHz is easy to achieve, but 30 Hz is a different story, especially with gear containing transformers. Harmonic distortion above 10 kHz matters less because the added harmonics are higher than the 20 kHz limit of most people’s hearing. However, if the distortion is high enough, audible IM difference frequencies below 20 kHz can result.
Sadly, many vendors publish only THD measured at 1 kHz, often at a level well below maximum output. This ignores that distortion in power amplifiers and gear containing transformers usually increases with rising output level and at lower frequencies.
The convention these days is to lump harmonic distortion, noise, and hum together into a single THD + Noise spec and express it as either a percentage or some number of dB below the device’s maximum output level.
For example, if an amplifier adds 1 percent distortion, that amount can be stated as 40 dB below the original signal. A-weighting is usually applied because it improves the measurement, and this is not unfair. There’s nothing wrong with combining noise and distortion into a single figure either when their sum is safely below the threshold of audibility.
But when distortion artifacts are loud enough to be audible, it can be useful to know their specific makeup. For example, artifacts at very low frequencies are less objectionable than those at higher frequencies, and harmonics added at frequencies around 2 to 3 kHz are especially noticeable compared to harmonics at other frequencies.
Again, this is why A-weighting is usually applied to noise and distortion measurements and why using weighting is not unreasonable.
1) The masking effect refers to the ear’s inability to hear a soft sound in the presence of a louder sound. For example, you won’t hear your wristwatch ticking at a loud rock concert, even if you hold it right next to your ear. Masking is strongest when both the loud and soft sounds contain similar frequencies.
As we have seen, the main reason to measure audio gear is to learn if a device’s quality is high enough to sound transparent.
All transparent devices by definition sound the same because they don’t change the sound enough to be noticed even when listening carefully.
But devices that add an audible amount of distortion can sound different, even when the total measured amount is the same. A-weighting helps relate what’s measured to what we hear, but some types of distortion are inherently more objectionable (or pleasing) than others.
For example, harmonic distortion is “musical,” whereas IM distortion is not. But what if you prefer the sound of audio gear that is intentionally colored?
In the 1960s, when I became interested in recording, ads for most gear in audio magazines touted their flat response and low distortion. Back then, before the advent of multilayer printed circuit boards, high-performance op-amps, and other electronic components, quality equipment was mostly handmade and very expensive. In those days design engineers did their best to minimize the distortion from analog tape, vacuum tubes, and transformers.
Indeed, many recordings made in the 1960s and 1970s still sound excellent even by today’s standards. But most audio gear is now mass-produced in Asia using modern manufacturing methods, and very high quality is available at prices even nonprofessionals can easily afford.
Many aspiring recording engineers today appreciate some of the great recordings from the mid-twentieth century. But when they are unable to make their own amateur efforts sound as good, they wrongly assume they need the same gear that was used back then.
Of course, the real reason so many old recordings sound wonderful is because they were made by very good recording engineers in great (often very large) studios having excellent acoustics. That some of those old recordings still sound so clear today is in spite of the poorer-quality recording gear available back then, not because of it!
Somewhere along the way, production techniques for popular music began incorporating intentional distortion and often extreme EQ as creative tools. Whereas in the past, gear vendors bragged about the flat response and low distortion of their products, in later years we started to see ads for gear claiming to possess a unique character, or color.
Some audio hardware and software plug-ins claim to possess a color similar to specific models of vintage gear used on famous old recordings. Understand that “color” is simply a skewed frequency response and/or added distortion; these are easy to achieve with either software or hardware, and in my opinion need not demand a premium price.
For example, distortion similar to that of vacuum tubes can be created using a few resistors and a diode, or a simple software algorithm.
The key point is that adding color in the form of distortion and EQ is proper and valuable when recording and mixing. During the creative process, anything goes, and if it sounds good, then it is good. But in a playback system the goal must be for transparency—whether a recording studio’s monitors or a consumer playback system.
In a studio setting the recording and mixing engineers need accurate monitoring to know how the recording really sounds, including any coloration they added intentionally. With a consumer playback system you want to hear exactly what the producers and mix engineers heard; you’ll hear their artistic intent only if your own system adds no further coloration of its own.
“The Audio Expert” by Ethan Winer, published by Focal Press (ISBN: 9780240821009), is available here. To read part 1, Audio Fidelity, Measurements, And Myths, go here.