Sample-rate conversion can produce different peak values because the converter filters and resamples the waveform rather than copying its original sample values. A 96 kHz master converted to 44.1 kHz is therefore a newly calculated signal, even when it sounds effectively unchanged.
You can see the practical problem when a master sits close to its ceiling. The source file may pass a true-peak check, yet the converted deliverable can land slightly higher because its output samples occupy different points along the reconstructed waveform. Peak safety measured only before conversion is incomplete.
This matters most when the source has been pushed hard. Dense limiting, clipped transients, and strong high-frequency content give the converter less room to perform its filtering without producing a higher crest somewhere in the output.
Filtering is the detail people often miss. A filter can reduce energy in part of the spectrum while still increasing a momentary peak because its phase and impulse response alter how frequency components line up in time. The converted waveform can therefore gain a taller instant without becoming meaningfully louder overall.
The issue is closely related to how dBFS and dBTP describe different peak behavior. Sample values tell you what exists on one digital grid, while true-peak measurement estimates the larger waveform those samples represent between grid points. Resampling builds another grid from that waveform.
A useful consequence follows. Two competent converters can produce files whose audible differences are tiny while their maximum samples are not identical. Filter design, passband behavior, phase response, and numerical implementation all influence the exact output, so a peak reading from the source is not a certificate for every conversion made from it.
Moving the converter before the final limiter changes the situation. The limiter then reacts to the destination-rate signal, which lets it control peaks created by the conversion itself. Such a chain can be sensible when one sample rate is the only required deliverable, although separate formats still deserve their own final checks.
Blindly lowering every limiter by several decibels is a crude fix. The safer margin depends on the material, the conversion path, and what processing follows. Nielsen and Lund documented overload risks in digital conversion in the Overload in Signal Conversion paper, including sample-rate conversion as a headroom-critical stage rather than a neutral file-format chore.
A cleaner workflow keeps the high-resolution parent master intact, creates the required sample-rate version, then measures the actual output file. If the converted file exceeds the ceiling you need, revise the gain or limiting at an appropriate point and render again. Do not assume the old meter reading still governs a changed signal.
Keep codec conversion separate in your head. Sample-rate conversion can alter peaks through filtering and resampling, while AAC, MP3, and other lossy encoders can produce additional changes later. A file that survives SRC cleanly can still peak differently after encoding, so each stage has its own failure point.
The same logic applies to batch exports. One song on an album may barely move after conversion while another gains enough peak level to cross a delivery ceiling. Program material matters, which makes a single project-wide assumption weaker than measuring every finished file.
A final render check should use the exact sample rate, bit depth, and processing order intended for delivery. Read true peak from that file, listen for obvious conversion problems, and keep enough headroom for any later stage you do not control. The useful number is the peak of the deliverable, not the reassuring number you saw one render earlier.
You can see the practical problem when a master sits close to its ceiling. The source file may pass a true-peak check, yet the converted deliverable can land slightly higher because its output samples occupy different points along the reconstructed waveform. Peak safety measured only before conversion is incomplete.
This matters most when the source has been pushed hard. Dense limiting, clipped transients, and strong high-frequency content give the converter less room to perform its filtering without producing a higher crest somewhere in the output.
Resampling creates a different set of sample values
A sample-rate converter first has to reconstruct enough of the signal to calculate samples for a different time grid. Downsampling also needs low-pass filtering so frequencies above the new Nyquist limit do not fold back into the audible band. Neither operation promises to preserve the numerical peaks in the source file.Filtering is the detail people often miss. A filter can reduce energy in part of the spectrum while still increasing a momentary peak because its phase and impulse response alter how frequency components line up in time. The converted waveform can therefore gain a taller instant without becoming meaningfully louder overall.
The issue is closely related to how dBFS and dBTP describe different peak behavior. Sample values tell you what exists on one digital grid, while true-peak measurement estimates the larger waveform those samples represent between grid points. Resampling builds another grid from that waveform.
A useful consequence follows. Two competent converters can produce files whose audible differences are tiny while their maximum samples are not identical. Filter design, passband behavior, phase response, and numerical implementation all influence the exact output, so a peak reading from the source is not a certificate for every conversion made from it.
The limiter ceiling can become stale after conversion
Suppose you limit a 96 kHz master and measure its final true peak before creating the 44.1 kHz release file. The limiter has controlled the signal at the point where it sits in your mastering chain. Sample-rate conversion comes later and performs another filtering operation, so its output is no longer the exact signal the limiter saw.Moving the converter before the final limiter changes the situation. The limiter then reacts to the destination-rate signal, which lets it control peaks created by the conversion itself. Such a chain can be sensible when one sample rate is the only required deliverable, although separate formats still deserve their own final checks.
Blindly lowering every limiter by several decibels is a crude fix. The safer margin depends on the material, the conversion path, and what processing follows. Nielsen and Lund documented overload risks in digital conversion in the Overload in Signal Conversion paper, including sample-rate conversion as a headroom-critical stage rather than a neutral file-format chore.
A cleaner workflow keeps the high-resolution parent master intact, creates the required sample-rate version, then measures the actual output file. If the converted file exceeds the ceiling you need, revise the gain or limiting at an appropriate point and render again. Do not assume the old meter reading still governs a changed signal.
The delivery-rate file deserves the final meter pass
True-peak measurement after conversion catches something a source-session meter cannot. It evaluates the samples that will actually leave your hands, including peak changes introduced by the resampling filter. For a 44.1 kHz delivery, meter the rendered 44.1 kHz file rather than treating the 96 kHz reading as interchangeable.Keep codec conversion separate in your head. Sample-rate conversion can alter peaks through filtering and resampling, while AAC, MP3, and other lossy encoders can produce additional changes later. A file that survives SRC cleanly can still peak differently after encoding, so each stage has its own failure point.
The same logic applies to batch exports. One song on an album may barely move after conversion while another gains enough peak level to cross a delivery ceiling. Program material matters, which makes a single project-wide assumption weaker than measuring every finished file.
A final render check should use the exact sample rate, bit depth, and processing order intended for delivery. Read true peak from that file, listen for obvious conversion problems, and keep enough headroom for any later stage you do not control. The useful number is the peak of the deliverable, not the reassuring number you saw one render earlier.
