A 24-bit file holds about 144.5 dB of dynamic range, while 32-bit float holds a theoretical 1528 dB and keeps everything above 0 dBFS intact.
That one format choice decides whether a peak problem is fixable or permanent. Clip a 24-bit fixed-point file, and the damage is written into the samples. Clip a 32-bit float file, and the peak is still sitting there, waiting to be turned down.
The same asymmetry runs through the whole handoff. Almost every peak decision costs less the earlier you make it, and the cost curve gets steep fast.
Pull the level down in the DAW, and that audio behaves like any other file, undistorted. In 24-bit fixed point, the same overshoot is gone. Gone for good, and no amount of downstream processing brings it back.
This is why the old advice about how much headroom your mix actually needs deserves retiring. The peaks-at-minus-six rule comes from the DAT era, when converters and tape machines punished hot input, and it survives mostly out of habit.
Two rules replaced it. Do not let peaks reach 0 dBFS in a fixed-point format, and do not print a limiter on the mix bus, because the second one hands the mastering engineer a ceiling they cannot remove.
Every one of those touches a single element and leaves everything else alone. That is the whole advantage, and it disappears the moment the mix is printed.
Stems narrow the gap without closing it. Grouping the mix into drums, bass, vocals, and effects gives a mastering engineer some element-level control, enough to compress one group harder or pull a harsh frequency out of one part without touching the others.
Abbey Road puts the limit plainly: stem mastering is not a mix fixing service. A stem is still a sum. If the peak lives inside the drum stem, you are still choosing between the peak and everything else in that stem.
With a stereo file, you are choosing for the entire record at once. Every move is global, which is why mastering decisions get argued over in tenths of a decibel while mix decisions get made in whole ones.
The useful dividing line is where the peak came from. A single unintended event, a chair creak, a fader bump, a stray transient buried in a loop, belongs to the mix, because one person can remove it in one edit without consequences.
Structural peaks are different. When the kick, the bass, the vocal, and the crash all land on the same downbeat, nothing is wrong with any of them individually. Every element is doing its job, and the sum still breaks the ceiling.
Those are the peaks that legitimately reach mastering, and on a delivered stereo file the appropriate response is envelope work applied to the sum rather than the source. You cannot unstack the arrangement, so you flatten the crest it produces and leave the rest of the song at full height.
Format choice quietly moves that boundary. Deliver 32-bit float, and a peak that ran past the mix bus ceiling is still recoverable at the mastering stage, so the mistake stops being permanent and becomes an inconvenience.
Deliver 24-bit, and the boundary is hard. Whatever clipped on the way to the file belongs to the mix engineer permanently, whether they noticed it or not.
That one format choice decides whether a peak problem is fixable or permanent. Clip a 24-bit fixed-point file, and the damage is written into the samples. Clip a 32-bit float file, and the peak is still sitting there, waiting to be turned down.
The same asymmetry runs through the whole handoff. Almost every peak decision costs less the earlier you make it, and the cost curve gets steep fast.
Bit depth decides whether a clip is recoverable
The numbers are not close. Both formats put 0 dBFS in the same place, but 32-bit float extends roughly 770 dB above it, which means signals over the ceiling are preserved in the file rather than flattened against it. Both figures are theoretical ceilings rather than anything a converter reaches, but the gap between them is the point.Pull the level down in the DAW, and that audio behaves like any other file, undistorted. In 24-bit fixed point, the same overshoot is gone. Gone for good, and no amount of downstream processing brings it back.
This is why the old advice about how much headroom your mix actually needs deserves retiring. The peaks-at-minus-six rule comes from the DAT era, when converters and tape machines punished hot input, and it survives mostly out of habit.
Two rules replaced it. Do not let peaks reach 0 dBFS in a fixed-point format, and do not print a limiter on the mix bus, because the second one hands the mastering engineer a ceiling they cannot remove.
At the mix, you have the source; at the master, you have the sum
A mix engineer looking at a rogue peak has options that simply do not exist later. Mute the doubled guitar hit that nobody notices. Drop the velocity on one crash. Shorten the sample. Edit the breath. None of that is available once the file is a stereo bounce.Every one of those touches a single element and leaves everything else alone. That is the whole advantage, and it disappears the moment the mix is printed.
Stems narrow the gap without closing it. Grouping the mix into drums, bass, vocals, and effects gives a mastering engineer some element-level control, enough to compress one group harder or pull a harsh frequency out of one part without touching the others.
Abbey Road puts the limit plainly: stem mastering is not a mix fixing service. A stem is still a sum. If the peak lives inside the drum stem, you are still choosing between the peak and everything else in that stem.
With a stereo file, you are choosing for the entire record at once. Every move is global, which is why mastering decisions get argued over in tenths of a decibel while mix decisions get made in whole ones.
Some peaks are supposed to reach mastering
None of this means a mix should arrive pre-flattened. Overprepared mixes are their own problem, and they turn up constantly, squeezed defensively until there is nothing left for anyone downstream to shape. A mastering engineer handed a flat file can make it louder and not much else.The useful dividing line is where the peak came from. A single unintended event, a chair creak, a fader bump, a stray transient buried in a loop, belongs to the mix, because one person can remove it in one edit without consequences.
Structural peaks are different. When the kick, the bass, the vocal, and the crash all land on the same downbeat, nothing is wrong with any of them individually. Every element is doing its job, and the sum still breaks the ceiling.
Those are the peaks that legitimately reach mastering, and on a delivered stereo file the appropriate response is envelope work applied to the sum rather than the source. You cannot unstack the arrangement, so you flatten the crest it produces and leave the rest of the song at full height.
Format choice quietly moves that boundary. Deliver 32-bit float, and a peak that ran past the mix bus ceiling is still recoverable at the mastering stage, so the mistake stops being permanent and becomes an inconvenience.
Deliver 24-bit, and the boundary is hard. Whatever clipped on the way to the file belongs to the mix engineer permanently, whether they noticed it or not.
