Normalization Basics
Normalization is a DSP calculation, not a very nasty one, but it adds a minute amount of (probably imperceptible) distortion. But a little bit of imperceptible distortion accumulates the more processing that the track goes through. If you are going to be sending your material for mastering, do NOT normalize. To repeat: THERE IS NO NEED TO NORMALIZE IF THE MATERIAL IS GOING TO BE FURTHER PROCESSED. Let the next DSP step or analog processing step take care of two birds with one stone, be in the hands of the mastering engineer to avoid additional calculations, etc. In general, normalization should be avoided. In my book I cover this in more detail, but basically, once a track has already been recorded, you do not gain any quality by changing its gain, you only lose quality by requantizing it. If you are mixing it, you are going to be changing the gain once again anyway, so why do an extra quality-reducing DSP step prior to mixing?
Normalization In-depth
Can Normalization improve the sound? This is an extremely complex issue, you have to examine all the variables.
“Bob, If you want to hear an improvement in broadband, take a 16 bit file, normalize and capture it to 24 bits. Whether or not you capture at higher bits, the improvement in relative LSB from the normalizing make a huge difference.”
George, thanks for the advice. I hope you don’t mind an essay on all the possible implications of your statement, in order to be very exact about it and not confuse any readers. There are a lot of implications which can happen due to potential misinterpretation of your statement. I’m concerned about your statement “whether or not you capture at higher bits,” for there is a serious loss if you do not use the longer wordlength that comes from any recalculation.
If Sonic’s desk and EDL did not calculate in 24-bits (or greater) accuracy, then Sonic Solutions’ excellent sound character would be compromised! Perhaps you are just referring to the benefits of normalization in terms of getting the signal above the “garbage level” of typical reproduction systems by the simple act of a gain increase. What I mean by “signal to garbage ratio” is the line noise; RFI injection, digital noise on the grounds, D/A converter noise; lack of monotonicity at low levels, and other problems normally present in any digital reproduction system, which decreases one’s enjoyment of the music. Please note that the *original* *signal to noise ratio* of the source 16-bit material is fixed and cannot be improved by the gain increase or act of normalization (ignoring special techniques like No-Noise). This is very important to realize. Only the ultimate “signal to garbage” ratio of the final reproduction system is improved.
By the way, a high signal-to-garbage ratio is very important, I have done matched level listening tests that show that raising the gain to arrive at a 0 dBFS peak can significantly improve the listening enjoyment of even a very good 20-bit reproduction system (you get greater improvement when raising gain with lower class home-type systems, which have more built-in “garbage”). But (and a big but) you have to consider the longer wordlengths that are generated.
I hope you’re not implying that “you can get something for nothing by, for example, changing the gain of a file, say .1 dB or so, then capturing as many of the extra bits as possible… Are you implying that the extra bits that result somehow contain new important information? The extra bits result from requantization… the extra bits that result actually contain the original information, but spread around a larger wordlength, and wherever you chop them off, some quantization distortion results. This, of course, requires redithering. Now we must evaluate the trade off in the improvement in signal-to-garbage ratio against the requirement of the addition of dither. If you take a 16-bit source, raise its gain 3 dB and then redither to 16 bits and truncate, the sound will probably be deteriorated, because the new dither adds a veil at -96 dBFS (for purposes of discussion); the original source’s dither is now raised to approximately -93 dBFS (the sum of two noise floors). That means you have two dither signals, including a new dither that’s only 3 dB below the original source’s dither and contributes to a sonic veil. This is a lose-lose situation.
But if you raise the gain of the 16 bit source 3 dB and redither to 24 bits for reproduction on a 24 bit DAC or 20-bit system, and never return to 16 bits, then the sound may improve, for the signal to garbage ratio has been increased. A 24-bit DAC is cleaner and quieter than a 16 bit DAC, and the sound may improve. But never forget that the original noise floor of the original source has now been raised by 3 dB. You get the improved signal to garbage ratio when you turn down your monitor gain that 3 dB (turning down the garbage). But you will only get the improvement if you both raise the gain and move up to a 24-bit reproduction system!
Dither is a tradeoff, since the dither noise itself acts as a veil on the very ambiance it is trying to preserve. “Dither, you can’t live with it, you can’t live without it!” You have to decide on a case by case basis if the tradeoff is worth going through the exercise of the normalization, or gain increase, requantization, and required redithering. Every recalculation is potentially a loss in resolution, not an improvement, due to questions of how much gain you are adding, versus the trade off of the sonic veiling that will inevitably arise from the addition of dither noise to requantize back to 16 bits at the end. 3 dB gain is probably not enough to consider on the basis of SNR alone, if you remain in 16-bits.
For example, if I had a well recorded 16-bit set of tracks that sounded excellent, had good loudness characteristic for the musical material being presented, however, it peaked to -1 dBFS, I would probably not choose to raise it 1 dB (normalize it). Because the compromise of the addition of dither would likely cause a subtle loss of transparency rather than an improvement in sound. This is a serious subjective judgment process that must be examined every step of the way; I cannot make a blanket statement that there would be an improvement, and to my ears, 16-bit—>processor—>16-bit is almost always an audible loss. You can minimize that loss by using the very best noise-shaped dither, until it is almost imperceptible in many musical cases.
Don’t forget to consider the precision of the gain calculation itself, which can add low-level distortion. This must be done with high resolution, ideally 48 bits; this necessitates calculating an extra long dither word as well.
Interestingly, recent listening studies of poor-quality (consumer) D/A converters show they like to see low level bits “exercised,” and it seems that these poor converters seem to like the addition of some dither noise as they improve in low-level accuracy.