Please see the expanded version of this article on the ProSoundWeb site.
The expanded version goes into much more depth, and names names.
(This page remains only to support existing links from other web sites.)

Conventional wisdom says that setting digital record levels to peak around -20 dB below full scale sounds better than recording at levels closer to the 0 dBFS (Full Scale) maximum. One popular web article claims that recording at lower levels avoids "recordings that sound 'weak' or 'small' or too dense or 'just not pro enough'" among other complaints. However, the specs for most sound cards and outboard digital converters show a similar frequency response and distortion amount for a wide range of input levels. So it doesn't make sense that recording at low levels should sound better or even different. Further, most modern DAW software uses 32-bit floating point (FP) math, so signal levels within the software have very little effect on sound quality. If recording at low levels really does sound better in a given setup - as proven by a proper level-matched blind test - it's more likely due to improper gain-staging elsewhere in the analog portion of the signal path.

The benefit of 32-bit floating point math is explained in detail in my book The Audio Expert, so I won't elaborate further here. The main point (no pun intended) is that using 32-bit FP math accommodates a range of signal levels exceeding 1,000 dB. Therefore, most plug-in effects sound exactly the same no matter what volume level you send through them. Now, some plug-ins are affected by signal level, particularly those that add intentional distortion or other "vintage analog" effects when overdriven. But most plug-ins accommodate the same huge range of levels as their DAW software host. There should be no audible or measurable difference between mixing a group of tracks that were each recorded near full scale, versus mixing tracks that are all 20 dB softer and raising the level later in the chain. As long as the master output volume is adjusted to avoid distortion in the rendered Wave file, both mixes will sound exactly the same.

In the article mentioned above, the author suggests creating an entire project with each source recorded simultaneously onto two tracks at levels 20 dB apart, then making parallel mixes to prove that recording at lower levels sounds wider, clearer, and generally better overall. But apparently he never actually did that test, or he would have realized he's wrong! The author also claims the improvement in audio quality is more apparent with a larger number of tracks due to the stacking effect, but that's a different issue and is also a myth. Of course, it's a good idea to keep levels low when recording a live concert, for safety to avoid distortion if something unexpected comes along. But it's easy to show that the recorded quality per se is not improved at lower levels.

To save everyone the bother, I did this test and the results confirm what I've been saying for a long time in audio forums. Since it's not practical for me to record an entire band in my home studio which requires dozens of microphones and simultaneous input channels, I did the next best thing: I played each track of an existing song one by one, and re-amped (re-recorded) them using a loudspeaker and microphone to two tracks at the same time with the levels 20 dB apart. Whether re-amping a loudspeaker captures the same sound as a microphone on a singer or drum set is irrelevant. The source simply is what it is, and a loudspeaker source is as valid as any other to disprove this myth. I used the 8-track "master" of the old Motown hit Ain't No Mountain High Enough that made the rounds a while ago.

The photo at left shows the setup with my large JBL 4430 loudspeaker and precision DPA 4090 microphone 18 inches in front of the center of the horn. The output of the mic preamp went through a Y-splitter, then into both inputs of my Focusrite sound card. I set the levels so that one track of each recording peaked at least above -6 dBFS, with the other 20 dB softer. I didn't change the record levels as each existing track was recorded to a new pair of tracks. I also recorded at 16 bits instead of 24 to make this a worse-case test. Since I already had a basic mix of these tracks, I simply solo'd each track as I recorded each copy pair. Both groups of copied tracks were sent to their own output bus whose volumes were set to peak just below 0 dB, then rendered to separate Wave files. I exported only 25 seconds of the tune to keep the file sizes reasonable (each 4.5 MB):