How it works

How Aupiphany levels a recording

The app takes the same six steps with every file, in this order. It opens with its No AI switch on, and with that on, a classic audio tool does every step. The pictures are drawn from recordings the engine leveled, the same ones that play on the home page. Under the steps is a list of every tool in the chain.

  1. Step 1

    It measures loudness the way the ITU standard defines it

    Aupiphany measures loudness with the method in ITU-R BS.1770, the standard behind every loudness target in podcasting and broadcast. The audio first passes through a filter that cuts the lowest frequencies and lifts the highest a little, roughly the way the ear weighs sound. The level is then averaged in 400 millisecond blocks. Blocks quieter than −70 LUFS, and blocks more than 10 LU under the average, are left out, so pauses do not pull the number down.

    The result is the integrated loudness of the whole file. The recording in the picture measured −20.6 LUFS before leveling.

    Integrated loudness
    −20.6LUFS
    The original recording of two readers, before anything changes.
  2. Step 2

    It finds speech, music and silence

    Every 40 milliseconds the engine looks at the sound and marks it as speech, music, noise or silence. It decides by measuring how the level moves, because speech rises and falls with the syllables and music holds steadier, so no trained model is involved. Speech is brought to one common level. Music is kept 1.5 LU under the speech, so a music bed stays under the voices instead of rising to meet them.

    Pauses take the gain of the speech around them, so room noise is not turned up on its own. With the Standard leveling setting, long pauses between phrases are also lowered by 6 dB.

    The first 30 seconds of a podcast episode, leveled. The strip along the top marks speech in gray and music in violet.
  3. Step 3

    It draws a slow gain curve that looks ahead

    Aupiphany reads the whole file before it changes anything, so it already knows what comes next. It finds the points where one speaker hands over to another by looking at the pauses: when the level on each side of a pause differs by more than 4 dB, it treats the new voice as a new speaker and corrects it from the first syllable.

    Inside one speaker, an 8-second window sets the gain, and a 2-second window takes over only when the two disagree by more than 3 dB, so natural emphasis stays. The gain moves in the pauses between words, which is why the result does not pump. Leveling boosts or cuts by up to 18 dB. The Standard setting applies 90 percent of the correction it works out, Light applies 60 percent, and Strong applies all of it.

    Gain applied, from −2.2 dB
    +16.7dB
    The gold line is high exactly where the original is short: that stretch is the quiet reader.
  4. Step 4

    It sets the whole file to the target

    Before leveling, a high-pass filter at 70 Hz takes out rumble under the voices. After leveling, a compressor rounds off the syllable peaks that are left, at a ratio of 2.2 to 1 with the Standard setting. Then the whole file is turned up or down by one fixed amount so its integrated loudness lands on the preset's target.

    The Podcast preset aims for −16 LUFS in a stereo file and −19 LUFS in a mono file, because a mono file at −19 plays as loud as a stereo file at −16.

    The two-reader recording, which is stereo, landed on −16.0 LUFS.

    Loudness, from −20.6
    −16.0LUFS
    The leveled recording hangs below the line. Both readers now reach about the same depth.
  5. Step 5

    It holds true peaks under the ceiling

    A file stores samples, and the waveform a player rebuilds between those samples can rise higher than any single sample. That higher point is the true peak. Aupiphany finds it by looking at the audio at four times the sample rate for files below 96 kHz, the method BS.1770 describes.

    Two stages keep true peaks under the ceiling, which is −1 dBTP for the Podcast preset. A fast peak compressor with a soft knee rounds off the sharpest peaks first, and a look-ahead limiter then catches whatever is left.

    MP3 and AAC encoders shift peaks and loudness a little as they compress. For those formats Aupiphany measures the shift on short test encodes first and takes it out of the final file. When the ceiling is left on automatic, it is lowered to −1.5 or −2 dBTP for low bit rates, where encoders overshoot the most.

    True peak
    −1.0dBTP
    The dashed line is the −1 dBTP ceiling. No peak in the leveled file crosses it.
  6. Step 6

    It measures the finished file again

    When the export is written, Aupiphany decodes the saved file and measures it with the same meter. If the encoder moved the loudness by more than 0.1 LU or pushed a peak over the ceiling, it renders the file again with the difference taken out, up to twice. The result screen then shows the numbers read from the exported file, which are the numbers of the file you will upload.

    Loudness
    −16.0LUFS
    True peak
    −1.0dBTP
    Loudness range, from 15.4
    4.2LU
    Measured by the engine on the leveled recording.

Every tool in the chain, in order

Aupiphany opens with its No AI switch on. With it on, the six steps above are carried out by the tools below, and none of them uses machine learning, so no trained model and no neural network touches your audio. Nothing is uploaded, and all of it runs on your Mac, iPad or iPhone. Broadcast and recording engineers have used most of these kinds of tools for decades: gain riding, compressors, peak limiters, loudness meters and noise reduction.

  1. Noise reduction Optional

    This stays off until you switch it on. It is made for steady noise such as hiss, hum, a fan or air conditioning. It measures how loud every moment of the recording is and takes the quietest tenth as the moments when no one is talking, because a voice pauses and a fan does not. The average sound of those moments is the noise.

    Each frequency band is then turned down by how close it sits to that noise, by at most 6 dB on Light, 10 dB on Medium and 15 dB on Strong.

  2. Low-cut filter

    A fourth-order filter at 70 Hz removes the rumble that sits under the voices, the kind a microphone picks up from a desk or from traffic outside.

  3. Speech and music detection

    Every 40 milliseconds the engine marks the sound as speech, music, noise or silence, using plain measurements of how the level moves over the second around that moment. Speech rises and falls with the syllables and leaves gaps between them. A stretch that holds steadier is marked as music, and no trained model takes part in the decision.

  4. Gain rider

    This is the automatic form of an engineer riding a fader. Every 10 milliseconds it works out how loud the voice at that moment is. It finds a change of speaker in the pauses, where the level on each side of a pause differs by more than 4 dB, so a quiet guest is corrected from the first syllable.

    It listens about 4 seconds ahead and 4 behind and moves the gain during pauses, which keeps the level steady inside each word.

  5. Two compressors

    The first compressor rounds off the syllable peaks that are left after the gain rider, at a ratio of 2.2 to 1 on the Standard setting. The second is a fast peak compressor with a soft knee, and it rounds off the sharpest peaks before they reach the limiter.

  6. True-peak limiter

    A look-ahead limiter holds true peaks under the ceiling, which is −1 dBTP on the Podcast preset. True peaks are measured at four times the sample rate for files below 96 kHz, the method ITU-R BS.1770 describes.

  7. Loudness meter

    The meter measures integrated loudness to ITU-R BS.1770 and EBU R 128, the broadcast loudness standards. The whole file is then turned up or down by one fixed amount so that reading lands on the preset's target.

  8. Silence trimming Optional

    This also stays off until you switch it on. It trims silence at the start and end of the exported sound file and shortens long pauses. Anything under a level threshold counts as silence. You can set that level yourself or leave it on automatic, which picks a level a little above the recording's own background noise.

  9. Read-back check

    After the export is written, Aupiphany decodes the saved file and measures it with the same meter. If the MP3 or AAC encoder moved the loudness by more than 0.1 LU or pushed a peak over the ceiling, it renders the file again with the difference taken out, up to twice.

The one AI option Off until you choose it

The No AI switch is in Adjust and in Settings, and it is on when the app opens. Turning it off adds one choice under Background noise, labeled “Isolate voice (uses AI)”, for recordings made in noisy rooms. It uses Apple's on-device machine learning to separate the voice from the room, and it takes the place of noise reduction at the start of the chain. It works on your device like the rest of the chain, so nothing is uploaded.

After every export the result screen says “No AI used on this file” when that is true. When Isolate voice was used, it says that voice isolation used Apple's machine learning.

Glossary

The four numbers Aupiphany reports, in plain terms.

LUFS
Loudness units relative to full scale: how loud a whole program sounds, measured with the ITU-R BS.1770 method. One LU is one decibel of loudness, so −16 LUFS is 2 LU louder than −18 LUFS. Apple Podcasts asks for about −16, and some standards write the same unit as LKFS.
True peak
The highest point of the waveform once it is rebuilt between samples, in decibels relative to full scale (dBTP). A ceiling of −1 dBTP leaves room for MP3 and AAC encoders, which can push peaks up when they compress a file.
Loudness range
How much the loudness moves over a program, in LU, defined in EBU Tech 3342. It is the gap between the 10th and the 95th percentile of the short-term loudness, measured in 3-second windows with very quiet passages left out.
Speech spread
Aupiphany's own figure for how far apart loud and quiet voices are: the gap between the 10th and the 90th percentile of the short-term loudness during speech, in LU. The two-reader recording went from 15.1 to 3.4.

Coming soon to the App Store for Mac, iPad and iPhone.

See pricing