Podcast loudness comes down to one number pair: about -16 LUFS integrated, with a true peak at or below -1 dB. That's what Apple publishes for podcasts, and it applies to AI voice audio exactly as it does to a microphone recording. Hit it and your episode plays at a similar level to the shows around it.
This post gives the targets with sources, a command that measures your file, a second command that fixes it, and a method for the case that actually causes trouble: an episode where AI-voiced segments sit next to recorded ones and one of them is suddenly louder.
What LUFS should a podcast be
LUFS (or LKFS, the same unit under another name) measures loudness the way ears hear it, averaged over a whole file. True peak measures the highest point the signal reaches after the audio is reconstructed, which can be above the highest sample you see in an editor. Both are defined in ITU-R BS.1770, and both matter for different reasons: loudness decides how the episode feels next to others, and true peak decides whether it distorts when it's compressed to MP3 or AAC.
As of October 2026, here is what the platforms say.
| Source | Integrated loudness | True peak | Applies to |
|---|---|---|---|
| Apple Podcasts | -16 LKFS, plus or minus 1 dB | At most -1 dB FS | Podcast audio |
| Spotify (loudness page) | -14 LUFS | Below -1 dB TP | Music tracks only |
The Apple figures come from its podcast audio requirements page (Apple Podcasts audio requirements), which also recommends embedding loudness metadata so playback engines can adjust levels. Spotify's published numbers are on its loudness normalization page (Spotify loudness normalization), and that page covers music.
I couldn't find a podcast-specific loudness target in Spotify for Creators' documentation, so I won't claim one.
My recommendation follows from that: master to -16 LUFS with a true-peak ceiling around -1.5 dB. The extra half decibel is my own margin for lossy encoding, not an Apple rule. It sits inside Apple's tolerance, and it doesn't leave you quiet next to anything else on a phone.
Why AI voice isn't exempt: A generated voice has no mic distance or room to give it a natural level, so it lands wherever the generation put it. It can come out quieter or louder than your recorded audio, and the platform can't tell the difference. Measure it like any other file.
How to measure podcast loudness with ffmpeg
ffmpeg is free and includes an EBU R128 meter. This command reads the file, prints a summary and writes nothing. Free alternatives with a loudness meter or normalizer include Audacity's Loudness Normalization effect, if you prefer a window to a terminal.
ffmpeg -hide_banner -nostats -i episode.mp3 -af ebur128=peak=true -f null -The summary lists Integrated loudness (I), Loudness range (LRA) and True peak (Peak). Compare I against -16 and Peak against -1. If I is -23 and your peak is comfortable, the episode is simply quiet and needs about 7 dB of gain. If I is right but the peak is -0.2, you have a limiting problem instead.
Normalize in two passes
ffmpeg's loudnorm filter can correct the file. Running it twice, once to measure and once to apply, preserves the dynamics better than a single pass. Take the four values from the first run's print_format=json output and feed them to the second.
# Pass 1: measure
ffmpeg -hide_banner -i episode.wav -af loudnorm=I=-16:TP=-1.5:LRA=11:print_format=json -f null -
# Pass 2: apply, using the measured_* values from pass 1
ffmpeg -i episode.wav -af loudnorm=I=-16:TP=-1.5:LRA=11:measured_I=-23.1:measured_TP=-4.2:measured_LRA=6.8:measured_thresh=-33.4:offset=0.5:linear=true -ar 44100 episode-16.wavKeep a lossless master (WAV) until the last step, then encode to MP3 or AAC once. Apple's page says to precondition the audio before encoding, so normalize first and compress second. VocalLab's podcast export gives you MP3 or WAV, with WAV available on Pro and above, so choosing WAV for the episode gives you a clean file to run through the commands above.
Export settings that survive the feed
Apple's RSS feed guidance accepts MP3 or AAC. For mono spoken word it suggests 64 to 128 kbps at 44.1 or 48 kHz, and for stereo 128 to 256 kbps. Speech is happy at the low end of those ranges. For a one-voice show, a mono file at 64 to 96 kbps is a reasonable way to keep a 40-minute episode small.
Two habits matter more than the bitrate. Export the same sample rate for every segment so nothing gets resampled mid-edit. And generate with the narration tool in one pass where you can: the AI podcast narration page takes a full episode script and renders one continuous track, which means fewer joins to level.
Match levels between AI voice and recorded segments
The target numbers are the easy part. The harder problem is an episode built from parts, say a recorded intro, an AI-voiced body and a recorded outro, where each sounds fine alone and jarring in sequence. The fix is to level the parts to each other before you level the whole.
Measure each part on its own
Run the ebur128 command on each segment. Ignore anything shorter than a few seconds, since integrated loudness is unreliable on very short clips.
Pick a reference
Choose the part with the steadiest level. For many shows that's the AI voice, since there's no mic distance to wander.
Move the others to within 1 LU
Apply gain to the other segments until their integrated loudness is within about 1 LU of the reference.
Normalize the whole episode
Now run the two-pass command on the assembled file to land on -16 LUFS and the true-peak ceiling.
Listen at the joins
Meters agree on averages, not on how a cut feels. Play ten seconds either side of every transition on phone speakers.
Also scan the short-term loudness in the ebur128 output, not just the summary. A single AI line that came out several decibels louder than its neighbors can pass the integrated check and still hurt on headphones. If one line is off, regenerate that line or lower it by hand; regenerating a single line costs only that line's points, at about 1 point per second.
Two voices that work as the AI half of a show, both from the library: a warm, natural female voice for a recurring host and a clear, measured male voice for explainer segments. Whichever you use, keep it the same from week to week. A voice that changes also changes the level you matched to.
Warm Natural Female Explainer Voice
A recurring host or narrated segments
- Neutral American
- Warm, natural
Explainer segments and long-form narration
- Neutral American
- Clear, measured
Mistakes that cost listeners
Normalizing to peak, not loudness
Peak normalization to 0 dB says nothing about loudness. Two files peaking at the same level can differ by many LU.
Skipping the final true-peak check
Encoding to MP3 after mastering can push peaks over the ceiling. Re-measure the encoded file once.
Mastering to a music target
The audiobook export's Streaming preset is -14 LUFS, 2 dB louder than Apple's -16. It's fine for audiobooks, but for a podcast feed you may prefer to export with no mastering and normalize yourself.
If you're building a recurring host, the podcast host voice page covers reusing one voice every week, which also keeps your levels steadier from episode to episode.
For casting options, see the AI podcast voice generator guide.
Frequently asked questions
What LUFS should a podcast be?
As of October 2026, Apple recommends around -16 LKFS (the same unit as LUFS) with a tolerance of plus or minus 1 dB, and a true peak no higher than -1 dB FS. Spotify's published -14 LUFS figure covers music, not podcasts.
How do I measure podcast loudness with ffmpeg?
Run ffmpeg with the ebur128 filter and peak=true on your file, with null output. The summary at the end gives integrated loudness in LUFS, loudness range and true peak, which you compare against -16 and -1.
Does AI voice audio need different loudness than recorded audio?
No. Platforms measure the finished file, not how it was made. The practical difference is that generated audio has no natural level, so measure it and match it to your recorded segments before mastering the whole episode.
Should I export my podcast as MP3 or WAV?
Export WAV when you can, so you can normalize on a lossless master and encode to MP3 or AAC once at the end. WAV export is available on Pro and above; MP3 plays everywhere.