Every voice at the same level.

No AI touches your audio unless you choose it.

Aupiphany evens out loud and quiet speakers, handles the music under them, and sets the finished file to the loudness your platform asks for.

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

Free to download. $29.99 once for the full version.

Both play at the same loudness

Loudness, from −20.6
−16.0LUFS
True peak
−1.0dBTP
Speech spread, from 15.1
3.4LU

A loud voice and a quiet one. Two volunteer readers from LibriVox (public domain), recorded separately and cut together at different levels. Leveled by Aupiphany. https://librivox.org

The numbers are the engine's own measurements of the whole recording this clip was cut from. For the comparison, the original is turned up 4.3 dB so it plays as loud as the leveled version.

Read what is said in this clip

Name, again answered Mr. Noun.

Do not try to be experienced in the presence of an audience. It may proceed from the gaze of the many eyes that turn upon the speaker, especially if he permits himself to steadily return that gaze. Most speakers have been conscious of this in a nameless thrill, a real something pervading the atmosphere, tangible, evescent, indescribable. All writers have borne testimony.

“Difel, sir,” said the judge sternly, “what is your name? Answer at once and truly.” “I have answered truly,” replied Mr. Noun. “My name is Name, for noun means name. The name of everything belongs to me.”

Transcribed from the leveled clip by speech recognition software and not yet checked against the recording by ear, so a word may be wrong.

Modern versions of the tools radio has run on for decades

Aupiphany opens with its No AI switch on. With it on, every step below is a classic audio tool with no machine learning in it, so no trained model and no neural network touches your audio. Your recording passes through them in this order, on your own device, and nothing is uploaded. Engineers in radio and recording studios will know most of them: gain riding, compressors, peak limiters, loudness meters and noise reduction.

  1. Noise reduction Optional

    Switch it on for a recording with hiss, hum or a fan in it. It measures that steady noise in the quietest tenth of the recording, then turns each frequency band down by how close it sits to the noise.

  2. Low-cut filter

    Takes out the rumble below 70 Hz that sits under the voices, the kind a microphone picks up from a desk or from traffic outside.

  3. Speech and music detection

    Tells speech from music by measuring how the level moves. Speech rises and falls with the syllables and leaves gaps, and a stretch that holds steadier is marked as music, with no trained model involved.

  4. Gain rider

    Does what an engineer does with a hand on the fader. It measures how loud each voice is every 10 milliseconds, finds a change of speaker in the pauses, listens about 4 seconds ahead and 4 behind, and moves the gain during pauses so no word changes level partway through.

  5. Two compressors

    The first rounds off the syllable peaks the gain rider leaves, at a ratio of 2.2 to 1. The second is faster, with a soft knee, and catches the sharpest peaks.

  6. True-peak limiter

    Looks ahead and holds every peak under the ceiling. It measures true peaks, the ones that fall between samples, the way ITU-R BS.1770 describes.

  7. Loudness meter

    Measures the whole file to ITU-R BS.1770 and EBU R 128, the standards broadcasters measure loudness by. The file is then set to your target from that reading.

  8. Silence trimming Optional

    Switch it on to trim the start and end and shorten long pauses in the exported file. A level threshold decides what counts as silence.

  9. Read-back check

    After export, Aupiphany decodes the finished file and measures it again. If the MP3 or AAC encoder moved the loudness or a peak, it corrects for that, so the numbers you see come from the saved file.

The one AI option Off until you choose it

Turning the No AI switch off adds one option, Isolate voice, for recordings made in noisy rooms. It uses Apple's on-device machine learning, so it also runs on your device and nothing is uploaded. Every export says “No AI used on this file” when that is true, and tells you when voice isolation was used instead.

Hear it on three common problems

Each clip is a real recording that Aupiphany leveled. Press play, then switch between Original and Leveled while it plays. The original is turned up to the same loudness as the leveled version, so you compare how even it sounds, not how loud.

A host and a remote guest

The host and the guest in this interview were not recorded at the same level. Across the whole episode, the gap between loud and quiet speech went from 6.1 LU to 2.6 LU.

Both play at the same loudness

Loudness
−19.5 to −16.0 LUFS
Loudness range
10.4 to 3.0 LU

Excerpt from “The R-Podcast Episode 13: Interview with Yihui Xie” by Eric Nantz, CC BY 3.0, via archive.org. Leveled by Aupiphany. https://archive.org/details/trp_audio_013

Original turned up 2.8 dB for this comparison.

Read what is said in this clip

viz.sivestat.com and you can find the repository on GitHub. Excellent. We'll put that in the show notes for sure. We'll put the link to that repository. What I want to do is to make it easier for authors to cooperate with each other using the GitHub model and also make it easier for reviewers to interact with authors. Because, you know,

This clip starts and ends in the middle of a sentence. The transcript was made by speech recognition software and has not yet been checked by ear.

Music between the voices

This is the opening of an episode: the title read aloud, a music sting, then the host. In the original the music is much quieter than the talking around it. Aupiphany marked the music, shaded violet in the drawing, and raised it to sit just under the speech instead of matching it.

Both play at the same loudness

Music Speech

Loudness
−20.2 to −16.0 LUFS
Loudness range
5.3 to 3.7 LU

Excerpt from “Going Linux #340: Security on Linux” by Larry Bushey, CC BY 4.0, via archive.org. Leveled by Aupiphany. https://archive.org/details/glp340

Original turned up 4.0 dB for this comparison.

Read what is said in this clip

Going Linux, Episode 340, Security on Linux.

[Music]

Welcome to the Going Linux Podcast. I'm your host, Larry Bushey. Whether you are new to Linux, upgrading from Windows to Linux, or just thinking about moving to Linux, this podcast will provide you with valuable information and advice that will help you in your future.

This clip is the start of the episode and ends in the middle of a sentence. The transcript was made by speech recognition software and has not yet been checked by ear.

An episode much quieter than the rest of a feed

This broadcast measured −36.3 LUFS. Played after an episode near −16, the usual podcast target, a listener would reach for the volume. Aupiphany brought the whole recording to −16.0 LUFS.

Both play at the same loudness

Loudness
−36.3 to −16.0 LUFS
Loudness range
19.4 to 6.6 LU

Excerpt from “KMUD (shortwave) Mojave Phone Booth Broadcast” by KMUD-sw, CC BY 3.0, via archive.org. Leveled by Aupiphany. https://archive.org/details/KmudshortwaveMojavePhoneBoothBroadcast

Original turned up 13.3 dB for this comparison.

Read what is said in this clip

Mojave phone booth. Hello? Anybody on the other end there? Nothing but a bunch of static from the lightning storms. Okay.

Much of this clip is radio static between the words. The transcript was made by speech recognition software and has not yet been checked by ear.

How to read the result

After Aupiphany levels a file, it shows the recording as three lanes on one timeline. This drawing comes from the same clip as the player at the top of the page.

Landed loudness, from −20.6
−16.0LUFS
  1. The original recording

    Drawn in lilac, growing upward. Tall where the loud reader speaks and short where the quiet one does.

  2. The gain Aupiphany applied

    The gold line rises where the original was quiet and dips where it was loud. In this clip it runs from −2.2 to +16.7 dB.

  3. The leveled recording

    Drawn in ivory, hanging downward like a reflection, so you can compare it with the original directly above.

  4. The landed loudness

    The integrated loudness of the finished file, with the starting number beside it.

The same leveler inside your editor

The Mac and iPad apps include an Audio Unit plugin, so you can measure and level inside the editor you already use. Put it on your editor's output and it meters the program against the preset's target. In the full version it also levels as the session plays.

A plugin hears the program as it plays, so it only knows the loudness of the whole thing after one full pass. It shows the running number, then offers the exact trim to the target. For a finished file, the app does the job in one step.

In the free version the meters work in full and leveling is switched off.

Where it runs

The plugin is an Audio Unit effect, installed with the Mac and iPad apps. It runs in editors on Mac and iPad that load Audio Unit effects, on a track or on the main output.

What it shows

  • Momentary and short-term loudness against the target
  • Integrated loudness so far and the loudness range
  • The highest true peak against the ceiling
  • The gain being applied at that moment

Presets, and where each number comes from

Podcast is selected when the app opens, so a new file needs no setup. Each preset sets a loudness target and a true peak ceiling.

PresetTargetPeak ceilingWhere the number comes from
Podcast−16 LUFS−19 LUFS for a mono file−1 dBTPApple Podcasts' published spec for −16. A mono file at −19 plays as loud as a stereo file at −16
Mono podcast−19 LUFS−1 dBTPFor a show made in mono. A mono file at −19 plays as loud as a stereo file at −16
YouTube and Spotify−14 LUFS−1 dBTPCommonly used. Spotify publishes −14 for music and YouTube publishes no number
Broadcast EBU R128−23 LUFS−1 dBTPThe European broadcast standard
ATSC A/85−24 LKFS−2 dBTPThe US television standard

Podcast, Mono podcast, and YouTube and Spotify use an automatic ceiling: −1 dBTP, lowered to −1.5 or −2 dBTP for MP3 and AAC files at low bit rates, where encoders overshoot the most.

Price

Free to download, and free to try on your own recordings.

$29.99 once for the full version on your Mac, iPad and iPhone.

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

What each version includes