Podcasts are getting longer; transcripts are getting essential
Edison Research's Infinite Dial 2026, released 12 March 2026, puts the figure at 58% of Americans age 12+ — 167 million people — who listened in the last month, and Pew Research notes that the average regular listener consumes multiple shows per week, with episode lengths trending upward — a 40-to-90-minute conversation is now normal where 25 minutes used to be the format ceiling. That is a lot of audio to keep in your head. A transcript is the cheapest way to make a long episode searchable, quotable, and re-readable.
Three steps, no account
- Copy the episode link. Open the show on Apple Podcasts or the podcast's own site and use "Share" / "Copy link" — or copy the show's RSS feed URL (we take the newest episode). RSS enclosure URLs (the direct mp3) work too.
- Paste it into shrybe. Our extraction pipeline pulls just the audio. It understands most major podcast platforms and any direct mp3 / m4a / wav link. Usually 5 to 20 seconds depending on episode length.
- Read, copy, or download. Whisper writes the transcript. You see timestamped paragraphs in your browser, ready to Copy or Download as a .txt. The whole job for a 30-minute episode typically completes in under a minute.
What podcasters and listeners actually do with transcripts
- Show notes and time-coded chapters. Most listening apps now display chapters and links pulled from show notes. A transcript with timestamps cuts hours of manual scrubbing — copy the segments that matter, drop them into your notes template.
- Repurposing into blog posts and newsletters. A 60-minute interview becomes a 1,500-word essay much faster when you start from the transcript rather than re-listen. Transcripts also let you quote your guest verbatim, which matters when the guest is carefully precise.
- Search inside episodes. Search is the killer feature broadcast podcasts never had. With a transcript saved locally, Cmd-F finds the moment in the episode where the guest mentioned a specific book, library, study, or city.
- Accessibility for Deaf and hard-of-hearing audiences. The WHO counts 430 million people who require rehabilitation for disabling hearing loss, and projects it will pass 700 million — one person in ten — by 2050. Audio-only podcasts exclude this group by default; a transcript published alongside the audio fixes that.
- Journalism and research. A quotable, time-coded transcript is the working unit of reporting. shrybe is not a courtroom-grade transcript service — for that you still want a human transcriber — but for early drafting, fact-checking, and "wait, what did they actually say?" moments, the speed is the point.
Should you use shrybe or Otter or Descript?
Otter.ai and Descript are excellent tools that are not what we are. The honest difference matrix:
- Use Otter.ai when you need real-time meeting transcription with speaker labels and a paid plan you'll log into every workday.
- Use Descript when you want to edit audio and video by editing text — that integrated studio is its whole point and it is worth the subscription if you publish a podcast.
- Use shrybe when you have an episode URL and you want a transcript right now without installing anything, signing up, or paying a per-minute fee.
These are not direct competitors. Otter and Descript are workflow tools you build habits around. shrybe is the equivalent of a microwave — you put a URL in, you get text out, and most days you don't think about it.
Languages, accents, and where Whisper struggles
Two numbers get confused here, so it's worth separating them. The 2022 paper (Radford et al.) describes a model trained on roughly 680,000 hours across 99 languages. The large-v3-turbo checkpoint running here was trained on more than five million hours and adds a hundredth language, Cantonese. And the human comparison people like to quote covers English only: 25 recordings, where Whisper landed within a fraction of a percentage point of professional human transcription (Figure 7). No equivalent comparison was published for any other language, so "near-human" is an English claim and nothing wider. Where it struggles: Ukrainian Surzhik on very short clips (Whisper's auto-detect frequently misreads it as Russian), heavy regional accents in low-resource languages, and very loud podcast intros that bury the first few seconds of speech.
We work around the Slavic auto-detect problem by pre-detecting the language from the
episode title, description, and uploader name before Whisper runs. If the metadata
text contains Ukrainian-only letters (і, ї, є,
ґ), the picker auto-selects "Ukrainian" with a "detected from source"
hint. You can override manually any time; your pick wins.
Privacy, retention, and what we keep
The audio is fetched from the podcast host to our backend, runs through Whisper, and the resulting text is delivered back to your browser over a live connection — including when that browser is Safari on a phone. We cache the audio for at most 24 hours so the job can be retried if your connection drops, then it is deleted by a nightly job. We never train a model on the free-tier traffic and we never share what passes through.
There is no account. The free tier rate-limits by IP at 15 transcripts per 24-hour window. We log only the IP, the source URL, and basic timing metadata, purely to enforce rate limits and debug failures. Those logs rotate out inside 30 days.
What shrybe actually reads out of a podcast feed
A podcast feed is RSS, and the audio hangs off exactly one element. The
RSS 2.0 specification
(version 2.0.11) makes <enclosure> an optional child of
<item> — but when it is there, all three of its attributes are
required:
url— where the file is. The spec's words: it "must be an http url".length— how big the file is, in bytes.type— a standard MIME type, typicallyaudio/mpeg.
Paste a feed URL and shrybe reads the newest item's enclosure. Paste an enclosure URL straight from the feed and it skips the feed entirely — both work. The audio is fetched from the source rather than from your device, so a 90-minute episode costs you a link and a page of text, not a download.
This also explains the one failure that looks like a bug and is not: an item with no
<enclosure> has no audio attached to it. Some shows publish
chapter-only or notes-only items, and a few platforms serve a landing page in place of
a feed. There is nothing to transcribe in either case, and shrybe says so rather than
guessing.
Frequently asked
Which podcast platforms does shrybe support?
Apple Podcasts episode links (the URL contains "?i="), any podcast RSS feed (we transcribe the newest episode), the enclosure mp3/m4a URLs inside feeds, SoundCloud tracks, and any direct audio link. Spotify is the one big exception — its streams are DRM-locked, so paste the same episode from Apple Podcasts or the show’s RSS instead. If you can download the file, we can transcribe it.
Can shrybe transcribe a private or members-only episode?
No. We never log into Patreon, Apple Subscriptions, or other paywall providers on your behalf. If you have a Patreon-exclusive episode you legitimately have access to, download it through the official Patreon app and upload the resulting file to shrybe (direct upload is free, up to 100 MB per file).
How long is the longest episode shrybe can handle?
The free tier caps each transcript at roughly 75 minutes of audio, which covers most full episodes. For a 90-minute interview, split it before pasting, or use the paid fast lane, which takes files up to three hours. The cap exists because Whisper compute is real cost, and a fully free service has to draw the line somewhere.
Do podcasters need our permission to transcribe their show?
Fair use generally covers personal transcription for note-taking, study, or accessibility. Re-publishing the transcript on a public site as your own content is a different conversation — that is a copyright matter between you and the show. shrybe is a transcription tool; what you do with the text is on you.
Will the transcript include speaker labels?
Not yet. Whisper itself does not perform speaker diarisation — separating "Speaker A" from "Speaker B" — out of the box. Diarisation models exist (pyannote.audio, Sonnox SpeakerKit) and we are evaluating them for the paid tier. For now you get one continuous transcript with timestamps.
How accurate is Whisper on accented English or mixed-language episodes?
Whisper's published English error rates depend far more on the recording than on the speaker: across the fourteen English datasets in Radford et al. (2022, Table 2), Large V2 spans 2.7% word error on clean read speech to 36.4% on far-field meeting audio, averaging 12.8%. A close-miked podcast sits near the good end of that range, which is why episodes transcribe better than conference recordings. Mixed-language interviews — say a Ukrainian guest on an English-language podcast — usually work, but Whisper will often pick the dominant language and transcribe the rest phonetically. If you know the primary language, pick it manually in the language picker for cleaner output.
How is shrybe different from Otter.ai or Descript for podcast work?
Otter and Descript are full SaaS products with accounts, paid tiers, and meeting-recording features. shrybe is a single-purpose web tool — paste a public URL, get a transcript. No login, no per-minute meter, no upsell. If you need an editing studio that re-renders audio when you delete a word, Descript is what you want. If you need a transcript to read or quote from, shrybe is faster.
Are my transcripts private?
Yes. We delete the cached audio and the transcript text within 24 hours of the job completing. There is no account, so there is nothing to link a transcript back to a person. We never sell or share what passes through the free tier.
Sources
Every figure on this page traces back to a primary source, not a secondary summary:
- Edison Research, The Infinite Dial 2026 — released 12 March 2026. 58% of Americans age 12+ (167 million) in the last month; 45% (130 million) in the last week.
- RSS 2.0 specification,
version 2.0.11 — the three required attributes of
<enclosure>. - Radford, Kim, Xu, Brockman, McLeavey, Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision (2022, arXiv:2212.04356) — 680,000 hours across 99 languages, and the English-only human comparison in Figure 7.
- openai/whisper-large-v3-turbo model card — the five-million-hour training mixture and the hundredth language.
- WHO, Deafness and hearing loss — fact sheet dated 3 March 2026.
Got an episode link handy?
Paste it into the box at the top of this page. If you also have a YouTube video or an audio file URL, those work too. Episodes in another language can come back translated into English.