Your samples never touch 0 dBFS. Your meters are green. You export the file, it gets uploaded to Spotify, transcoded to AAC, and played back through a consumer DAC. Somewhere in that chain, the audio clips.
This is not a bug. It is a direct consequence of how digital audio reconstruction works, and it is called an intersample peak.
What Band-Limited Reconstruction Actually Does
A digital audio file stores a sequence of discrete sample values. At 48 kHz, there is one measurement every 20.8 microseconds. The samples themselves do not produce sound. What produces sound is the analog waveform reconstructed from those samples by a DAC, using a low-pass reconstruction filter.
This reconstruction filter is not simply connecting dots with straight lines. It interpolates a smooth, band-limited curve through the sample points, following the mathematics of the Nyquist-Shannon sampling theorem. The theorem guarantees that a band-limited signal sampled above the Nyquist frequency can be perfectly reconstructed. But "perfect reconstruction" describes the continuous waveform between the samples, not the sample values themselves.
That continuous waveform can reach amplitudes that none of the samples reach.
Consider a signal with two adjacent samples at, say, 0.95 and 0.93 normalized amplitude. The interpolated curve between them does not simply move from 0.95 straight down to 0.93. Depending on the frequency content of the signal, it can arc upward briefly, past 0.95, before descending. The resulting intersample peak might land at 1.02: two percent above digital full scale, invisible to any sample-domain meter.
This is not a pathological edge case. For complex, high-frequency signals near full scale, intersample peaks of 1 to 3 dB above the highest sample value are common. Pop and EDM masters are frequently full of them, particularly after heavy limiting has pushed sample peaks close to 0 dBFS.
How ITU-R BS.1770 Measures True Peak
Sample-domain peak meters cannot detect intersample peaks by definition: they only read sample values. True-peak measurement solves this by simulating the reconstruction process before measuring.
ITU-R BS.1770, currently at revision 5 (November 2023), specifies the reference algorithm. For a 48 kHz signal, the procedure is:
- Upsample by 4x, producing a 192 kHz stream.
- Apply a 48-tap FIR low-pass filter during upsampling (12 taps per polyphase branch).
- Measure the peak absolute value of the resulting high-resolution stream.
- Report the result in dB TP (dB true peak).
At 192 kHz, there are four sample points for every original 48 kHz sample. The interpolated samples approximate the analog reconstruction curve, making previously invisible peaks visible. The standard specifies 4x as the minimum oversampling ratio and notes that implementations using higher ratios will achieve better accuracy. Some commercial true-peak meters use 8x or 16x.
The standard also specifies a 12.04 dB attenuation before the oversampling stage, applied to prevent integer overflow in fixed-point implementations. Floating-point systems can ignore this detail, but it appears in the specification for historical reasons.
A sample-domain peak meter reading -0.3 dBFS on a typical dense mix can easily correspond to a true peak of +0.4 dBTP or higher. That gap is precisely where codec-induced distortion lives.
Why Codecs Make It Worse
Intersample peaks already exist in the original PCM audio. Lossy codec encoding magnifies the problem.
When a streaming platform transcodes a WAV to AAC or Ogg Vorbis, the encoder modifies frequency content through quantization and psychoacoustic modeling. These modifications alter the waveform, and the altered waveform can reconstruct to even higher analog peaks than the original. A master delivered at -1 dBTP before encoding may produce true peaks above 0 dBTP after AAC encoding, particularly in the high-frequency range where codec processing is most aggressive.
This is why -1 dBTP exists as a mastering ceiling: it provides roughly 1 dB of headroom for codec processing to operate without introducing clipping on playback. Most streaming platforms have converged on this requirement. Spotify, Apple Music, Tidal and YouTube all specify -1 dBTP as the maximum true peak for uploads. Amazon Music specifies -2 dBTP, the strictest current requirement. Mastering to -2 dBTP satisfies all platforms with a single file.
The Look-Ahead Limiter: Catching the Future
A standard brickwall limiter reacts to the signal after it arrives. By definition, it cannot prevent an intersample peak that exceeds the sample values it is given. A look-ahead limiter changes this by giving the gain reduction stage advance notice.
The mechanism has four steps:
Step 1: Create a lookahead copy. The incoming signal is split. One path feeds a delay buffer. The other path is read some samples ahead of the delay output, providing a preview of what is coming. Typical lookahead times range from 1 to 10 ms. Five milliseconds is common: long enough to allow smooth gain curve computation, short enough not to introduce perceptible DAW latency during tracking (240 samples at 48 kHz).
Step 2: Detect the peak envelope. The preview signal passes through a peak-hold followed by a smooth envelope follower. The peak-hold prevents the envelope from missing short transients; the follower produces a smooth gain control signal rather than a jagged one.
Step 3: Compute gain reduction. For each sample in the preview, compute the gain factor required to bring the output below the threshold. Because the limiter knows the peak is coming, it can begin reducing gain before the peak arrives, spreading the gain reduction over several milliseconds. This eliminates the hard-knee distortion that would result from instantaneous gain reduction applied at the exact moment of a transient.
Step 4: Apply the gain to the delayed audio. The delayed audio path reaches the gain multiplication stage at the same moment as the corresponding gain reduction value. Because the timing is aligned, the gain reduction is applied precisely when and where it is needed.
A true-peak-aware limiter runs this same algorithm on the oversampled representation of the signal. It measures peaks in the 192 kHz domain, computes gain reduction there, and applies that reduction back to the original sample-rate audio. The result is a signal whose reconstructed analog waveform does not exceed the ceiling, not just a signal whose samples do not exceed it.
Practical Consequences
Sample peak meters are insufficient for mastering. Any loudness metering plugin used in a mastering context must include true-peak measurement. STC, LUFS, and sample peak are not substitutes for dBTP.
The safe mastering ceiling is -2 dBTP. At -1 dBTP, there is a non-trivial probability of post-codec true peaks above 0 dBTP on platforms like YouTube, where AAC encoding at lower bitrates is more aggressive. At -2 dBTP, the margin is sufficient for all current platforms and their transcoding pipelines.
Higher-frequency content generates larger intersample peaks. A mix heavy in high hats, distorted guitars, and synthesized transients produces larger intersample peaks than a mix dominated by bass and midrange. This is why dense electronic and metal masters are most vulnerable: the combination of heavy limiting and high-frequency content creates large deviations between sample peaks and true peaks.
Oversampling in your limiter matters. A limiter that operates only at the native sample rate cannot measure or control intersample peaks. When evaluating limiters for mastering work, check whether true-peak mode is available and what oversampling ratio it uses. Four times is the standard minimum. Eight times is better for signals with significant energy near the Nyquist frequency.
The Takeaway
Digital metering was designed to prevent sample overflow. It was not designed to represent the analog behavior of the reconstructed waveform. For most mixing applications, this distinction does not matter. For anything that will be delivered to a streaming platform, transcoded, and played through a DAC, it matters a great deal.
A true-peak meter running BS.1770-compliant 4x oversampling shows you what a DAC will actually produce. A look-ahead limiter controlling that measurement protects against the one form of clipping that sample-domain tools cannot see. The combination is the minimum technical standard for any contemporary mastering chain.