Reply to thread

Message: [QUOTE="Queen, post: 90619, member: 27"] ITU-R BS.1770-5 requires a true-peak meter to run at four times oversampling, and even at that rate the standard admits an under-read of roughly 0.55 dB. A sample-peak meter reads the numbers in the file and nothing else. Your converter draws a smooth curve between those numbers, and that curve can rise well above the highest sample it passes through. How big the gap gets depends entirely on the material. In extreme cases, inter-sample peaks can sit as much as 6 dB above the sample peak, which means a file your DAW swears is under the ceiling can still drive a converter into hard clipping. [HEADING=2]Sample peak counts numbers; true peak predicts voltage[/HEADING] The BS.1770 true-peak algorithm is not complicated. It attenuates the signal by 12.04 dB, upsamples four times to 192 kHz, low-pass filters, takes the absolute value, then adds the 12.04 dB back. The attenuation exists so the intermediate stages have somewhere to go without overflowing. The oversampling exists because you cannot see between samples without inventing more of them. Nothing exotic about either step. Four times is a compromise, and the standard says so. Its own table puts the worst-case under-read at four times oversampling at about 0.554 dB, so a meter reading exactly -1.0 dBTP may really be sitting near -0.45. None of this is visible on the meter your DAW shows by default, which is a sample-peak meter with a red light on it. It tells you whether any number in the file hit full scale. It has no opinion about what happens after the converter. [HEADING=2]The encoder is where the damage actually happens[/HEADING] Streaming platforms do not send your WAV file to listeners. They re-encode it, and lossy encoding rebuilds the waveform from a frequency-domain approximation, which can land higher than the source it started from. The bitrate matters here too, since a lower bitrate means a coarser approximation and a bigger potential overshoot. That is why the published numbers exist. Spotify normalizes to -14 LUFS and asks for masters below -1 dBTP, tightening that to -2 dBTP for masters louder than the target, because louder tracks suffer more encoding distortion. The AES lands in the same place. Its streaming recommendation, TD1008, states that maximum true peak should not exceed -1 dBTP at the codec input, for all content, without exception for genre or loudness target. The practical consequence is ugly. [B]Distortion that only appears after encoding[/B] is invisible in your session, invisible on your reference monitors, and audible on the version everybody actually hears. [HEADING=2]A limiter ceiling is not the same as delivered headroom[/HEADING] Set your limiter to a true-peak ceiling of -1, and you have not guaranteed -1 on delivery. You have guaranteed that one particular detector, running one particular oversampling ratio, could not see anything above -1. That is a narrower promise than it sounds. Different meters disagree for exactly this reason. Two true-peak readouts on the same file can land on different numbers, because they oversample at different ratios, and the under-read shrinks as that ratio climbs. The reliable move is to stop leaning on the ceiling. Hard limiting flattens the waveform, flat tops generate high-frequency content that was not there before, and sharp edges are exactly what produces large inter-sample overshoots. So the deepest gain reduction in your song is also the part most likely to overshoot after encoding. Reducing how deep the limiter has to go fixes both problems at once, and it does not require touching the ceiling at all. This is the case for [B][URL='https://goldmidi.com/community/threads/peak-tamer-reins-in-mastering-spikes-with-gain-automation.76018/']a peak-conditioning pass over the stereo file[/URL][/B] before the limiter sees it, so the loudest few moments arrive already level with everything around them. The limiter then works evenly instead of in bursts. Leaving margin is the other half. If your delivery target is -1 dBTP, limiting to -1.0 leaves you nothing, and the standard's own under-read figure says your meter may be optimistic by half a decibel. A ceiling of -1.5 or -2 costs you almost nothing after normalization, because Spotify is turning the track down to -14 LUFS anyway. Loudness normalization removed most of the reason to fight for that last decibel, and the fight is where the distortion comes from. The check worth adding is on the encoded file rather than the master. Encode a copy to AAC or Ogg Vorbis at the bitrate your distributor uses, decode it, and measure that. The number you get there is the one listeners are actually exposed to. [/QUOTE]

Name