An excerpt from Mastering Audio: The Art and The Science by Bob Katz.
May 10, 2012, by Bob Katz
This article is the second part in a series on decibels, excerpted from Bob Katz’s book Mastering Audio: The Art and The Science. The first part is available here, and be sure to check out all our videos featuring Bob Katz.
We Won’t Get Fooled Again. Recording engineers rely heavily on their favorite meter, and this is not intended to change people’s favorite.
But as practicing engineers, it is prudent to learn the defects and virtues of each meter we encounter.
The VU Meter. Relative newcomers to the industry may have never seen a VU meter, and some of them may be using the word “VU” incorrectly to describe peak-reading digital meters.
VU should only be applied to a true VU meter that meets a certain standard. The first thing we must learn is that the VU meter is a dreadful liar… It is an averaging meter, and so it cannot indicate true peaks, nor can it protect us from overload.
However the VU does do one thing better than a peak meter—it comes closer to our perception of loudness, but even so, it is a very inaccurate loudness meter because its frequency response gives low frequency information equal weight, and the ear responds less to low frequencies.
Another problem is that the VU meter’s scale is so non-linear that inexperienced operators think that the greater part of the musical action should live between -6 and +3 VU, but this is wrong.
A well-engineered music program has plenty of meaningful life down to about -20 VU, but since the needle hardly moves at that level, it scares the operator into thinking the level is too low.
Only highly-processed (dynamically compressed) music can swing in such a narrow range; in other words, the VU scale encourages over-compression.
Hence the VU meter should only be taken as a guide. A much better averaging meter would have a linear-decibel scale, where each decibel has equal division and weight down to -20 dB.
Digital Peak Meters
Digital Peak meters come in three varieties:
1. Cheap and dirty
2. Sample-accurate and sample-counting (but misleading)
3. Reconstruction (oversampling)
Cheap and Dirty Peak Meters. Recorder manufacturers pack a lot in a little box, often compromising on meter design to cut production costs. A few machines even have meters which are driven from analog circuitry—a definite source of inaccuracy.
VU meter operators are often fooled into treating the top and bottom halves of the scale with equal weight, but the top half has only 6 dB of the total dynamic range.
Some manufacturers who drive their meters digitally (by the values of the sample numbers) cut costs by putting large gaps on the meter scale (avoiding expensive illuminated segments).
The result is that there may be a -3 and a 0 dB point, with a large unhelpful no man’s land in between. When recording with a meter that has a wide gap between -3 and 0, it is best practice to stay well below full scale.
Sample-Accurate and Sample-Counting Meters. Several manufacturers have produced sample-accurate meters with 1 dB (or smaller) steps, that convert the numeric value of the samples to a representation of the sample value, expressed in
The Paradox of the Digital OVER. When it comes to playback, a meter cannot tell the difference between a level of 0 dBFS (FS = Full Scale) and an OVER. That’s because once the digital signal has been recorded, the sample level cannot exceed full scale, as in this figure.
We need a means of knowing if the ADC is being overloaded during recording. So we can use an early-warning indicator—an analog level sensor prior to A/D conversion—which causes the OVER indicator to illuminate if the analog level is greater than the voltage equivalent to 0 dBFS.
If the analog record level is not reduced, then a maximum level of 0 dB will be recorded for the duration of the overload, producing a distorted square wave.
After the signal has been recorded, a standard sample-accurate meter cannot distinguish between full scale and any part of the signal that had gone over during recording, it shows the highest level as 0 dBFS. However, a sample-counting meter can analyze a recording to see if the ADC had been overdriven.
This meter counts contiguous samples and can actually distinguish between 0 dBFS and an OVER after the recording has been made! The sample-counting digital meter determines an OVER by counting the number of samples in a row at 0 dB.
If 3 contiguous samples equal 0 dBFS, the meter signals an OVER, because it’s fair to assume that the incoming analog audio level must have exceeded 0 dBFS somewhere between sample number one and three.
Three samples at 44.1 kHz is a very conservative standard; on that basis, the recorded distortion would last only 33 microseconds and would probably be inaudible.
While an original analog signal can exceed the amplitude of 0 dB, after conversion there will be no level above 0, yielding a distorted square wave. This diagram shows a positive-going signal, but the same is true on the negative-going end.
While this type of meter was sophisticated in its day, current thinking is that the sample-counting meter is only suitable for evaluating whether an ADC has overloaded.
Authorities now feel that meters which display the digital value of the samples and which count samples to determine an OVER are no longer sufficient for mastering purposes and should be used with caution during mixing. Their place is taken by…
The Reconstruction Meter: Even More Sophisticated As long as a signal remains in the digital domain, the sample level of the digital stream is sufficient to tell us if we have an OVER. However, signals which migrate between domains can exceed 0 dBFS and cause distortion.
This includes any signal that passes through a DAC, a sample rate converter, or is converted through a codec such as mp3 or AC3. During the conversion from PCM digital to analog or mp3, filtering within the converter yields occasional peaks between the samples that are higher than the digital stream’s measured level, which can be higher than full scale.
This next figure shows that contrary to what we might assume, filtering or dips in an equalizer which we’d imagine would produce a lower output can actually produce a higher output level than the source signal. B.J. Buchalter explains that:
“the third harmonic is out of phase with the fundamental at the peak values of the fundamental, so it serves to reduce the overall amplitude of the composite signal.”
“By introducing the filter, you have removed this canceling effect between the two harmonics, and as a result the signal amplitude increases. Another reason for the phenomenon is that all filters resonate, and generally speaking, the sharper the filter, the greater the resonance.”
Equipment designers have known for years that because of filtering, the analog output level of complex audio from a DAC can exceed the sinewave value of 0 dBFS but very few have taken this into account in the design.
TC Electronic has performed tests on typical consumer DACs, showing that many of them distort severely since their digital filters and analog output stages do not have the headroom to accommodate output levels which exceed 0 dBFS!
While typical 0 dBFS+ peaks do not exceed +0.3 dBFS, some very rare 0 dBFS+ peaks may exceed full scale by as much as 4 or 5 dB with certain types of signals— especially mastered material which has been highly processed, clipped (turned into a square wave on top and bottom), and/or brightly equalized.
By oversampling the signal, we can measure peaks that would occur after filtering. An oversampling meter (or reconstruction meter) calculates what these peaks would be, but these meters are still rare. Products from TC Electronic (System 6000) and Sony (Oxford) have an oversampling limiter and reconstruction peak meter. RME’s Digicheck software includes an oversampling meter.
Reconstruction meters tell us not only how our DAC will react, but what may happen to the signal after it is converted to mp3 or sent to broadcast, both of which employ many filters and post-processes. Many DSP-based consumer players cannot handle the high levels at all and exhibit severe distortion with 0 dBFS+ signals.
Armed with this knowledge, no mastering engineer should produce a master that may sound acceptable in the control room but which she knows will likely produce severe distortion when post-processed or auditioned in the real world.
If the reconstruction meter is not enough to convince the client, she should also demonstrate that this “loud” signal becomes distorted, ugly, and soft when it is converted to low bit rate mp3. All the harmonics which made the signal seem loud in the control room have been converted to additional distortion.
Practice Safe Levels
What this means is that if you are mixing with a standard digital meter, keep peaks below -3 dBFS, especially if you are using aggressive bus processing.
The more severely processed, equalized or compressed a master, the more problems it can cause when it leaves the mastering studio.
We didn’t start hearing about this problem, or at least the severity of it, before the loudness race and the invention of digital processing which could be egregiously abused. Maximizing engineers should try to use a reconstruction meter and/or an oversampled brickwall limiter. If these are not available, use a standard peak limiter whose ceiling is set to -0.3 dB (see Chapter 10) and exercise caution.
But even the oversampled brickwall limiter is not foolproof; I’ve discovered that such limiters do not protect from very severe processing and can still make a consumer DAC overload unpleasantly. The best solution is to be conservative on levels. Clipping of any type is to be avoided, as demonstrated in Appendix 1.
The Myth of the Magic Clip Removal
If the level is turned down by as little as 0.1 dB, then a recording which may be full of OVERs will no longer measure any overs.
But this does not get rid of the clipping or the distortion, it merely prevents it from triggering the meter.
Some mastering engineers deliberately clip the signal severely, and then drop the level slightly, so that the meters will not show any OVERs.
This practice, known as SHRED, produces very fatiguing (and potentially boringly similar) recordings.
Peak Level Practice for Good 24-bit Recording
Even though 24-bit recording is now the norm, some engineers retain the habit of trying to hit the top of the meters, which is totally unnecessary as illustrated at left.
Note that a 16-bit recording fits entirely in the bottom 91 dB of the 24-bit. You would have to lower the peak level of a 24-bit recording by 48 dB to yield an effective 16-bit recording! There is a lot of room at the bottom, so you won’t lose any dynamic range if you peak to -3 dBFS or even as low as -10 dBFS, and you’ll end up with a cleaner recording.
Since distortion accumulates, if a “hot” recording arrives for mastering, the mastering engineer doing analog processing may have to attenuate the level to prevent the processing DAC from overloading. A digital mix that peaks to -3 dBFS or lower makes it easier to equalize and otherwise process without needing an extra stage of attenuation in the mastering.
In black is a complex wave. When the high frequency information (light orange) is filtered out, the result is a signal (orange) that is higher in amplitude than the original!
A number of 24-bit ADCs are advertised as having additional headroom, achieved by employing a built-in compressor at the top of the scale, claiming that the compressor can also protect the ADC from accidental overloads. But this is specious advertising.
Level accidents don’t occur in a mix studio; engineers have control over their levels and when tracking live musicians, it is better to turn off the ADC’s compressor, drop the level and leave plenty of headroom for peaks. The only possible use of this function of a compressor is if you like its sonic qualities and are trying to emulate the sound of tracking to analog tape.
But since tracking decisions are not reversible, I suggest postponing “analog simulation” to the mixing stage. It’s easier to add warmth later than try to take away some mushiness due to an overdriven compressor. As we have just seen, there is no audible improvement in SNR by maximizing a 24-bit recording and no SNR advantage to compressing levels with a good 24-bit ADC.
How Loud is It?
Contrary to popular belief, the levels on a digital peak meter have (almost) nothing to do with loudness.
Here is an illustration. Suppose you are doing a direct to two-track recording (some engineers do still work that way!) and you’ve found the perfect mix.
Leaving the faders alone, you let the musicians do a couple of great takes. During take one, the performance reached -4 dB on the meter; and in take two, it reached 0 dB for a brief moment during a snare drum hit.
Does that mean that take two is louder? No: because in general, the ear responds to average levels, not peak levels when judging loudness.
If you raise the master gain of take one by 4 dB so that it too reaches 0 dBFS peak, it will sound 4 dB louder than take two, even though they both now measure the same on the peak meter.
An analog tape and digital recording of the same source peaked to full scale sound very different in terms of loudness. If we make an analog tape recording and a digital recording of the same music, and then dub the analog recording to digital, peaking at the same peak level as the digital recording, the analog dub will have about 6 dB more intrinsic loudness than the all-digital recording.
Quite a difference! This is because the peak-to-average ratio of an analog recording can be as much as 12-14 dB, compared with as much as 20 dB for an uncompressed digital recording.
Analog tape’s built-in compressor is a means of getting recordings to sound louder (oops, did I just reveal a secret?). That’s why pop producers who record digitally may have to compress or limit to compete with the loudness of their analog counterparts.
The Myths of Normalization
The Esthetic Myth. Digital audio editing programs have a feature called Peak Normalization, a semi-automatic method of adjusting levels.
The engineer selects all the songs on the album, and the computer grinds away, searching for the highest peak level on the album and then automatically adjusts the level of all the material until the highest peak reaches 0 dBFS. If all the material is group-normalized at once, this is not a serious esthetic problem, as long as all the songs have been raised or lowered by the same amount.
But it is also possible to select each song and normalize it individually, but this is a big mistake; since the ear responds to average levels, and normalization measures peak levels, the result can totally distort musical values. A ballad with little crest factor will be disproportionately increased and so will end up louder than a rock piece with lots of percussion!
The Technical Myth. It’s also a myth that normalization improves the sound quality of a recording; it can only degrade it. Technically speaking, normalization adds one more degrading calculation and level of quantization distortion.
And since the material has already been mixed, it has already been quantized, which predetermines its signal-to-noise ratio—which cannot then be further improved by raising it.
Let me repeat: raising the level of the material will not alter its inherent signal-to-noise ratio but will add more quantization distortion. Of course material to be mastered does not need normalizing since the mastering engineer will be performing further processing anyway. Clients often ask: “do you normalize?” I reply that I never use the computer’s automatic method, but rather songs are leveled by ear.
A 24-bit recording would have to be lowered in level by 48 dB in order to reduce it to the SNR of 16-bit. The noise floors shown are with flat dither.
This is another form of normalization, an attempt to create an intelligent loudness algorithm based on the average level of the music, as opposed to the peak.
But when making an album, neither peak nor average normalization nor any intelligent loudness algorithm can do the right job, because the computer does not know that the ballad is supposed to sound soft.
There’s no substitute for the human ear. However, average normalization or better, a true intelligent loudness algorithm can help in situations where every program needs the same loudness, even if that doesn’t sound natural, such as radio broadcast, ceiling loudspeakers in a store, a party or background listening.
Judging Loudness the Right Way
Since the ear is the only judge of loudness, is there any objective way to determine how loud your CD will sound? The first key is to use a single DAC to reproduce all your digital sources and maintain a fixed setting on your monitor gain.
That way you can compare your CD in the making against other CDs, in the digital domain. Judge DVDs, CDs, workstations, and digital processors through this single converter.