We measured the most recent episode of 9,086 lively English-language podcasts. That is what we discovered. The total dataset is linked on the finish, and you’ll rating your personal present towards it in about twenty seconds.
13 p.c
The podcast statistics that flow into yearly say there are 4.58 million podcasts, 619 million listeners and a $39.6 billion enterprise. None of them say how the podcasts sound. No one had measured it.
So we did. Between September 1 and three we took the latest episode from 9,139 lively English-language podcasts, drawn at random throughout 19 classes, and ran three and three-quarter minutes of every via the identical perceptual mannequin and loudness meters that examine each file we ship. The strategy is on the finish.
13 p.c of podcasts clear the three fundamentals: loudness contained in the vary each distributor recommends, peaks under the extent at which shopper gadgets distort, and a perceptual rating above the road we deal with as acceptable for supply. 13 level zero p.c move all three. Eighty-seven p.c fail at the least one.
The fascinating half is which one.
Not the microphone
The same old story about dangerous podcast audio is a narrative about noise: a budget microphone, the untreated room. On that measure the medium is in affordable well being. The median episode scores 3.67 on a five-point perceptual scale (a broadcast studio sits close to 4.2, a nasty telephone line close to 2.5). Solely 11.7 p.c fall under 3.0, the purpose at which our personal pipeline flags a file. Voice readability, scored individually, has a median of 4.04.
The noise downside is usually solved. Microphones acquired low-cost and good, and the software program that cleans a room acquired cheaper and higher. The failures are elsewhere, within the a part of the chain no microphone touches.
Too quiet
Apple asks for minus 16 LUFS. Spotify normalises to minus 14. The targets have been public because the mid-2010s.
Of 9,086 episodes measured, 71.6 p.c sit exterior the minus 16 plus or minus 2 LUFS band. The median episode is at minus 19.7 LUFS, almost 4 dB beneath Apple’s goal and nearly 6 dB beneath Spotify’s. Two thirds (65.9 p.c) are under minus 18. Solely 5.7 p.c are too loud. This isn’t the loudness battle that flattened the report enterprise within the Nineties. It’s the reverse. Podcasts are made quietly, by individuals who by no means touched a fader, and shipped that manner.
The listener reaches for the quantity. Then the subsequent present within the queue, or a pre-roll inserted at a correctly normalised degree, arrives 6 dB louder.
Peaks, and a file twice the scale it must be
Two extra failures, each low-cost to repair and each widespread.
A 3rd of episodes (33.5 p.c) have true peaks above minus 1 dBTP, the ceiling under which a lossy encoder and a telephone speaker keep clear. Peaks above it are usually not all the time audible. They’re all the time avoidable, with a limiter that takes a second to use.
Two thirds of episodes (66.3 p.c) are stereo information. In 93.8 p.c of them the content material is a single centred voice, so the 2 channels carry the identical audio twice. The file is doubled in measurement, or halved in high quality on the similar bitrate, to encode a distinction that doesn’t exist. This can be a default no one modified.
For completeness, 10.6 p.c are encoded under 96 kbps and 5.9 p.c under 44.1 kHz.
Apply doesn’t make excellent
We anticipated reveals to enhance with age. They don’t. Exhibits with 5 to 25 episodes fail the perceptual threshold 12.3 p.c of the time; reveals with greater than 500 episodes, 16.1 p.c, the worst of the 4 tiers. Loudness is as flat. No matter a podcaster learns over time, it isn’t audio engineering, and nothing within the medium teaches it. The obtain counter doesn’t report LUFS.
The platforms
Internet hosting platforms are the one level within the chain the place a repair could possibly be utilized to tens of millions of information without delay, and several other promote one.
The information are detached to the advertising and marketing. Among the many massive hosts, the share of episodes under the perceptual threshold is bunched between 9.6 and 12.3 p.c: Buzzsprout 9.6, Podbean 10.5, Libsyn 10.6, Megaphone 11.9, Anchor and Spotify 12.1, Captivate 12.3. Transistor does greatest at 6.9. Spreaker (19.5), RedCircle (19.3) and SoundCloud (19.8) do worst. Loudness medians sit inside a decibel or two of each other, all wanting the goal. Platform is a proxy for who chooses it, not for what it does to the file. The platform with the mastering upsell is indistinguishable from those with out.
The place the difficulty concentrates
Sports activities podcasts are the cleanest, 4.4 p.c under threshold. Know-how 5.6, science 6.8, well being 7.0, enterprise 8.2. On the different finish, society and tradition 16.4 p.c and comedy 20.1 p.c, one in 5, which we suspect displays rooms with mates moderately than studios with engineers. Fiction, at 28.1 p.c, carries a caveat: audio drama is constructed from music and results, and a mannequin skilled on speech scores these as defects. That quantity is the mannequin’s restrict as a lot because the style’s.
What it means
A medium of 4.58 million titles has no loudness customary it enforces, no suggestions loop that teaches its producers one, and an 87 p.c failure fee on three measures a reliable engineer fixes within the time it takes to export the file. The noise downside everybody writes about is, by the numbers, the smallest of the 4.
What a podcaster can do about their very own present begins with figuring out the place it stands. That’s now measurable.
How we measured it
Body. The Podcast Index database export dated 2026-09-01 (MIT licence, podcastindex.org/datasets), 4,724,036 feeds. We stored feeds with an episode within the prior 90 days, at the least 5 episodes, a language tag starting en, a newest-episode period between 5 minutes and 4 hours and a major class: 229,066 lively English reveals. We took the 19 largest major classes (excluding music, whose episodes are music moderately than speech) and drew 500 feeds from every at random with a set seed, 139 from video games, which has solely 139. 9,139 feeds.
Fetch. For every feed, the most recent episode’s enclosure as listed within the export. 5 home windows of 45 seconds at 10, 30, 50, 70 and 90 p.c of the episode’s period, by no means inside the primary minute, by HTTP vary request at native pattern fee and channel rely. 9,086 episodes (99.4 p.c) yielded at the least three home windows. The 53 that didn’t had been eliminated (404), had been access-controlled (401, 403) or got here from hosts too sluggish to ship.
Measures. Perceptual high quality from Microsoft’s DNSMOS P.835 mannequin (SIG for speech, BAK for background, OVR general) on a 16 kHz mono by-product of every window, reported because the median throughout home windows, which tolerates a window touchdown on music or an commercial. Built-in loudness, loudness vary and true peak from the EBU R128 meter in FFmpeg over the concatenated native-rate home windows. Codec, bitrate, pattern fee and channel rely from the container. Loudness is subsequently measured on 3 minutes 45 seconds of programme audio per episode, a proxy for the full-episode worth; on our personal catalogue the 2 agree inside a decibel.
Thresholds. “Acceptable” is OVR at or above 3.0, the extent under which our pipeline flags a file. The loudness band is minus 16 plus or minus 2 LUFS. The height ceiling is minus 1 dBTP. “Clears the fundamentals” means all three without delay.
Limits. One episode per present. Class is self-declared within the feed. Fiction is scored by a speech mannequin. Cells beneath 100 reveals are omitted from the platform dialogue. Present measurement is by episode rely, as a result of obtain counts are usually not public.
Knowledge. The aggregates, together with sorted per-category distributions, are printed as JSON beside this report. Barevalue is a podcast manufacturing service; we measured the market we promote into and we are saying so. No present is known as.
Rating your personal present
Paste your feed at barevalue.com/rating and see the place your newest episode sits towards these numbers, with the three issues value fixing.
Knowledge
Pattern drawn from the Podcast Index open database (MIT). Fashions: Microsoft DNSMOS P.835, EBU R128 (FFmpeg). Snapshot 2026-09-01, scored 2026-09-03.
