Text: Jan Muths
The words: ‘as loud as possible without clipping’ brings up a respectable 15,000 results on Google.
From Europe to Australia, in books and online, among amateurs and professionals alike, this gain structure approach has embedded itself in the collective memory of the digital recording industry worldwide. But where did it come from and is it still the right digital recording method to employ today? Let’s look at the two most commonly stated reasons behind it and investigate the truth of the claim:
1: ‘Only when recording as loud as possible does one actually use all available bits’. This is almost true.
As you may already know, each bit represents around 6dB of the dynamic range, therefore the last bit is only used if the signal peaks between 0 and -6dBFS. This is a little simplistic, but we’ll discuss the details of this idea a little later.
2: ‘Every decibel of unused headroom reduces the signal-to-error ratio’. This statement is a fact – no doubt about it – but what it fails to take into account is that digital audio has evolved a great deal since studios swapped their beloved analogue tape machines for first-generation 16-bit recorders. Moreover, it’s now around 10 years since 24-bit converters squeezed their 16-bit counterparts out of the market, so it’s probably high time we had another critical look at the facts behind the above statements…
FROM 16 TO 24
There is wide acceptance in the audio industry that with higher bit resolution comes better sound, but it’s for a different reason than some may think. The step up from 16-bit to 20-bit doesn’t necessarily make our digital audio recordings any more aesthetically pleasing to the ear – at least not in the way that the switch from an entry-level compressor to a Neve 2254 might. Instead, this extra bit depth simply results in a lower quantisation error, not beautiful colouration. Of course, there’s also a big difference between entry-level converters and high-end A/Ds, which inevitably leads us to conversations about wordclock accuracy, jitter and other factors. But that’s a different story, one that has the potential to hijack this conversation and distract us from the issue under the microscope here. Today, let’s specifically look at bit depth.
Bit depth represents the amplitude of a signal and this is limited by two factors: the maximum level (defined as 0dBFS) and the minimum level, defined by the level of the quantisation error. The ‘distance’ between these two limits defines the dynamic range of a digital waveform. Between them the digital audio amplitude is scaled into small quantisation steps. The more of these a digital waveform possesses, the more accurately the amplitude-value can be captured. As we know from watching TV, the higher the pixel-resolution the more detailed and sharper the picture becomes. And the similarities continue. Just like the pixel resolution of a Full-HD TV, the number of quantisation steps is limited to a particular number. There cannot be an amplitude value between the quantisation steps, any more than there can be picture information between the pixels. Consequently, an A/D converter must either round the amplitude up or down to the nearest step in the grid. This process is called quantisation. The rounded (or truncated) value generates the quantisation error, essentially the equivalent of ‘noise floor’ in the analogue domain. Without question, this floor sounds worse than analogue noise of equivalent level.
Let’s have a look at how bit depth changes the level of quantisation error.
ERRORISATION
Figures 1 and 2 show an incoming 100Hz sine wave signal converted into a digital waveform using both 4-bit resolution (left) and 5-bit resolution (right):
The quantisation error can now be ‘isolated’ (as seen in Fig.3) by summing each signal from Fig.2 with the phase-inverted source signal. Notice that the 5-bit converter has a noticeably lower quantisation error level than the 4-bit equivalent.
The range between 0dBFS and the quantisation error level – or signal-to-error ratio – can be calculated as follows:
Signal-to-Error [SER] = (n x 6.02) + 1.76dB; where the variable ‘n’ equals the amount of bits we’re recording with. Recording at 16-bit thus establishes the SER as 98.08dB, while 24-bit converters establish a window of 146.24dB. In truth we should also quickly mention at this point that these figures represent the theoretical maximum at which 16- and 24-bit converters can ideally work. In the real world, and for several reasons we won’t delve into here, converters always provide a couple of dB less.
As mentioned earlier, as a rule of thumb each additional bit adds around 6dB to the dynamic range of a digital signal. Let’s ponder this fact for a moment: even 16-bit converters have a respectable dynamic range of almost 100dB, but how much dynamic range does a typical music CD actually possess? Not much by today’s brutal loudness standards. 100dB is far in excess of the dynamic range typically rendered to a mastered CD. The issues associated with dynamic range are really only relevant when multitracking is involved, particularly where the bit depth is the same as the destination format (ie: an album is multitracked at 16-bit and mastered to a 16-bit audio CD).
Let’s discuss an example in more detail.
DOUBLE TROUBLE
Let’s say we’re mixing down multiple tracks of a piece of music, each with the quantisation error level at –98dBFS (rounded down to a whole number to make the maths simpler). Every time we double the number of tracks, the quantisation error increases by 6dB – and that adds up fast!
The step from one to two tracks will increase the combined quanisation error by 6dB. So with a two-track mix our SER decreases from about 98dB to 92dB – still pretty good. But let’s double the track count again shall we? Now we’re working with only 86dB for four tracks. I’m sure you can see the pattern emerging here: as we double the track count we reduce our SER in steps of 6dB – 80dB for eight tracks, 74dB for 16, 68dB for 32 and so on. To be fair, we should mention that the quanisation error may not add up by exactly 6dB. It may be less depending on the error’s correlation.
Worse still, if we leave ourselves a couple of decibels of ‘headroom’ (which isn’t very much) to prevent us from straying ‘over’ the digital precipice, these must also be subtracted from our signal-to-error value, lifting the quantisation error value even higher. And we’re not even using compressors yet, which reduce dynamic range and effectively make soft signals louder still.
So, when we record at 16-bit we really need to keep the quantisation error low by recording as ‘hot’ as possible. This is of course how the philosophy: ‘Record as loud as possible without clipping’ or RALAPWC developed – a workflow derived from 16-bit conversion.
This fact alone has lead most of us to higher-than-16-bit resolution multitrack recording, even when the final product’s format is anticipated to be 16-bit. Although high-quality 16-bit converters still have their place in mastering and in two-track recorders, in 2011 they should be avoided for multitracking.
ABOUT HOW ‘HOT’?
At this point in the discussion there may now be even more readers feeling the need to convert audio from the analogue domain to digital as ‘hot’ as possible, to use more bits and keep the error value low. And if more bits means less error, we should all use higher-bit converters, right? In fact, why not use 32-bit converters, or even higher? 64-bit processing is currently being implemented in software processing, so why shouldn’t we abandon 24-bit converters for 64-bit A/Ds? All modern digital mixers (consoles and DAWs) already use higher than 24-bit processing internally after all. But while this is certainly useful for digital summing and processing, it offers no advantage in terms of conversion. Bear in mind too that the human ear has a range of around 120dB from threshold of hearing to threshold of pain. The dynamic range of a 24-bit converter, with its 16.7 million quantisation steps, is theoretically 146.24dB! I say ‘theoretically’ because in practise the noise floor of the analogue circuits inside a converter’s line inputs and outputs effectively limits the dynamic range. The analogue front-end actually has a higher noise-floor than a 24-bit converter.
In truth, the best converters typically offer a usable dynamic range of somewhere between 118 and 125dB. It is therefore impossible to use all 24 bits in any meaningful way let alone 32 or 64. In reality, only 20 to 21 bits of the 24 are available to us. The lowest of these are rendered useless, covered up by the omnipresent analogue noise floor. This is why 32-bit converters, though theoretically possible, couldn’t offer us any practical advantage since all these extra bits would only disappear under the noise floor. Long story short: 24-bit is more than enough in terms of conversion.
Nowadays, and with 24-bit resolution being the norm, the analogue domain is what defines the noise level. It accumulates when signal is passing through microphones, preamps, dynamic processors, EQs, channel strips and leads. In reality, for a 24-bit recording, the so-called ‘nasty sounding’ quantisation error is masked by analogue noise (unless one makes serious gain-structure mistakes!). What this means in practice is that 24-bit recording allows us to handle input levels in a much more relaxed way than the outmoded RALAPWC dogma suggests, and provides leeway for more headroom than many of us typically allow for.
GAIN THE ADVANTAGE
Let’s look at gain structure for a moment. We know from experience that 0dB (or +4dBu) is pretty much the optimal level in the analogue domain. 0dB makes sure our signal travels to the tape machine and back with the minimal introduction of tape hiss, it allows signals to travel long distances with minimal interference and keeps signals clean when routing music through daisy-chained units. We should also have a couple of decibels of headroom up our sleeves in case the talent gets a bit too excited. At ‘0dB’ on a VU meter, the signal sits well above the noise floor and a bit of headroom keeps our signal below the ‘red’ (where distortion starts to creep into the picture). Noise floor levels and headroom differ from unit to unit, but all professional devices work well at 0dB. If you stick with 0dB/+4dBu you should never have a problem with noise or distortion.
ZEROS AIN’T ZEROS
So how does this relate to the digital domain? Well, this is where some people get very confused. A major misunderstanding I’m often confronted with is the relation that exists between meters on an analogue console and the meters inside DAWs like ProTools etc.
First up, we must establish one simple fact: 0dB on an analogue meter is not the equivalent of 0dBFS in the digital domain! It never has been and never is. AT has covered this fact on several occasions in past issues, most recently in Andy Stewart’s editorial back in Issue 74.
So, how many dBFS is a dBu? This is not so easy to answer because one is measured in voltage, while the other is a digital value. In reality it depends on the converter itself and how it has been calibrated.
Unfortunately, there is no international standard for this relationship, although there are a few accepted norms, the most notable among them perhaps being in the film industry where 0dB = –20dBFS. In a calibrated system, this results in an analogue signal being represented on a VU meter at ‘0dB’, and –20dBFS inside the DAW.
‘What the heck?’ I hear you ask! ‘We just gave away 20dB! Don’t we get in trouble with the quantisation error if we do that, especially when we start summing heaps of tracks together?’
No, we don’t – this is the beauty of 24-bit: Our signal level at this point is still 126.24dB above the quantisation error! While the –98dBFS quantisation error of 16-bit converters sits above the analogue noise floor introduced by decent analogue front end – potentially causing headaches in the mix, particularly in soft passages of orchestral music, fade-outs or decay tails – our 24-bit quantisation error sits well below the analogue noise floor.
If you’re still not convinced, let’s briefly look at two popular devices – the Yamaha 02R96 and the new Avid HD I/O – as examples of how calibration works.
The Yamaha 02R96 operating manual comes with a detailed level diagram featuring dBu, dBFS and bit scale. Yamaha has chosen to calibrate +4dBu at –20dBFS. Therefore, if we maintain gain structure of +4dBu across our signal flow chain and route this signal digitally into ProTools, there would be little in the way of visible level – level only in the ‘green’ area, barely halfway up the meters! No need to worry though, this is not a problem if you’re recording at 24-bit.
Avid’s HD I/O operating manual states a similar level: +4dBu is set to –18dBFS as a factory default, but these interfaces can also be calibrated to different levels via their rear-mounted calibration screws – giving them the ability to interact with devices calibrated at higher or lower levels.
Recording ‘hotter’ than +4dBu (1.23V RMS) is not wrong of course, but always keep in mind that above this figure you’re exceeding the standard operating level and driving signals into the realm of analogue ‘headroom’. If that’s what you want to do, great! It’s entirely your decision. Just be aware that the harder you drive your analogue signal above the standard operating level, the more distortion you’re adding to the recording signal.
An example of this would be to record a signal using the Avid HD I/O with a final digital peak level of –6dBFS. In this case the analogue front end requires signal to be gained up to +16dBu (or 4.887V RMS respectively) – a level that not every analogue device can handle! SSL consoles may have enough headroom, but don’t try this with the original SPL Goldmike (unless you want lots and lots of distortion).
With today’s 24-bit converters offering so much room within which to move, we now have the choice. Record hot, record warm or even cool levels… it’s up to you. Just remember, regardless of what ‘temperature’ you choose, this decision should be based on your taste for the sound of the analogue front end and its noise-floor/THD, not the digital domain and its quantisation noise.
WHO CARES ABOUT RALAPWC?
Today, there’s nothing wrong with dropping the ‘RALAPWC’ dogma and following the gain structure of the good ol’ tape machine days.
You may want to prove this to yourself by testing your own recording gear: track a long guitar chord or a crash cymbal and listen carefully to the end of the decay tail. What do you hear: digital quantisation errors or the analogue noise floor? Now try the same recording again with levels 10 or 20dB under your previous to-tape levels. Chances are you’ll learn more about the noise floor of your analogue front-end doing this, while at the same time discovering that your modern converters will almost certainly handle both recordings equally well. This should give you a good indication of how ‘hot’ to record and how much headroom to leave yourself.
It’s time to let go of the RALAPWC workflow, as the problems that led us to this theory mostly no longer exist. Test your gear – you may well end up with recording levels around +4dBU – back to the good old days!
RESPONSES