The short version
shrybe is a free web tool that turns any public YouTube, podcast, or audio URL into a timestamped transcript using OpenAI Whisper. The free tier gives 15 transcripts per IP address per day, up to 75 minutes per file, with no account required and no ads on the page or in the output.
There is also a paid fast lane, and it sells exactly one thing: time. The free lane runs on our own machines at roughly real time, so an hour of audio takes about an hour and may queue behind someone else. The fast lane runs the same audio off-box in a fraction of that, never queues, and accepts files up to three hours. The model is identical on both, so paying gets you the same transcript sooner — not a better one. The free tier stays exactly as it is.
Why we built it
We watch a lot of technical videos for research — conference talks, deep-dives, benchmark walkthroughs, documentation streams. For a long time we treated pausing every sixty or ninety seconds to type out a quote or a command as the cost of consuming that kind of content.
That pause-and-type loop is not the cost of consuming technical content. It is friction that piles up until the typing is what you are doing, and the talk is just the thing you are trying to keep up with. Transcribe it once, then read or search — that part is obvious. What took longer to fix was the tooling. Tactiq requires a Chrome extension on every device. Otter.ai and Descript were built for meeting recordings and podcast editing suites — the account setup, the paid tiers, the feature surface — none of it fits the case where you have a public URL and want text in under a minute.
We had a working speech-recognition pipeline from prior product work. Moving it into a standalone transcription service took a weekend. Making it accept public YouTube URLs, podcast feeds, and direct audio links without requiring an account took a few more. The rate limits are real — fifteen transcripts per IP address per day, seventy-five minutes per file — and they exist because the compute we have for Whisper has a real ceiling and we want the free tier to stay free.
shrybe exists because we needed it. It is still the tool we use, several times a week.
About the small team behind this
There is a small team behind shrybe. We keep our personal identities out of the brand: no founder names on the front page, no LinkedIn cult of personality, no podcast tour, no recurring Twitter thread about the lessons from shipping startup #7. Reach us at hello@shrybe.com when you need a human; otherwise the tool is what we want you thinking about.
A surprising amount of SaaS about-pages are spent convincing you to trust the people who built the thing — credentials, prior exits, school logos, an inevitable photo of the team looking determined against a brick wall. We would rather let the thing speak. A transcript that works speaks for itself. One that doesn't — our names would not fix that.
What you get instead is the public /editorial-standards page that documents how we source, fact-check, and correct everything we publish; a clearly stated pricing model at /pricing; a public bug-report channel; and rate limits documented in plain numbers. That is the trust contract we are offering. We think it works.
How it works
The marketing site you are reading is static HTML. Every page is pre-rendered; there is no client-side framework loading in your browser. You get the page, and that is the page.
The transcription backend is a different story. When you submit a URL, our extraction pipeline pulls only the audio track (we never download or store the video itself), then feeds the audio to OpenAI Whisper — specifically the large-v3-turbo checkpoint, which keeps most of large-v3's accuracy at a fraction of its compute. The transcript returns to your browser live, streamed back as Whisper finishes, so the result still arrives even when the job runs longer than most CDN edge timeouts would tolerate.
Whisper itself was trained on roughly 680,000 hours of multilingual audio across 99 languages. Its published English error rates vary far more by recording conditions than any single headline number admits: across the fourteen English datasets in the paper's own comparison, the Large V2 model ranges from 2.7% word error on clean read audiobooks to 36.4% on far-field meeting microphones, averaging 12.8% (Radford et al., 2022, arXiv:2212.04356, Table 2). A clean interview lands near the good end of that range; a room with an open window does not. The turbo checkpoint gives up a little accuracy on accents and low-resource languages in exchange for substantially lower compute; the far lighter "small" checkpoint, which we run only behind the iOS Shortcut where a timeout forces our hand, gives up considerably more. The paid fast lane, live since August 2026, runs this same checkpoint: it buys speed, not accuracy. We unpack how that accuracy varies, language by language, in how accurate Whisper actually is.
Everything is small on purpose. There is no Kubernetes cluster, no auto-scaling group, no fleet. One backend, one team. When we outgrow that, we will reach for a managed Whisper API. Not before.
What we keep, what we don't, and what we don't hide
Three things shape what we do with the audio and transcripts that flow through shrybe.
What we keep: the audio file is cached on our server for at most 24 hours so we can retry if your connection drops; after that, a nightly job deletes it. The transcript text itself is delivered to your browser and not retained on our side. We log the IP address that submitted the request and the source URL purely to enforce the daily rate limit and debug failures — these logs rotate out within 30 days.
What we don't do: no advertising on the free tier (nothing pinkie-promise about it; the tier just was not designed with ads in mind). No data sales. No training of any model — ours or anyone else's — on what passes through the free tier. No third-party advertising pixels in the page HTML. We do use Microsoft Clarity and Google Analytics 4 on the marketing pages to understand which pages help and which don't, and the payment provider's pixel to see how many visitors who reach checkout complete it. All three load only after you click Allow in the consent note, and none of them receives the audio you submit or the transcripts we return.
What we don't hide: the editorial process at /editorial-standards, the pricing roadmap at /pricing, the deliberate-narrowness explanation in the section below, and any correction we ever have to issue (the policy is in editorial-standards too).
Two operating principles
Two principles shape what shrybe ships. Worth naming them because they explain some decisions that might otherwise look like stubbornness.
Positioning matters more than features do. A tool that calls itself a "video downloader" hits one search classifier; a transcription service hits a different one entirely, even when the underlying audio pipeline is the same. shrybe is positioned as transcription. The download capability in our backend stays below the fold; the headline is always "transcribe."
Fake trust signals are not neutral. shrybe ships with no
AggregateRating Schema.org block, no testimonial carousel, no synthetic
"join 10,000 happy customers" counter, no fabricated quote, and no statistic that does
not link to a primary source we verified ourselves. Every number on this site has a
citation. Every quote exists in the cited source. Google Quality Raters, AI search
engines, and your own scepticism all reward this — but the honest reason we do it is
that we find sites which don't to be exhausting to use.
Why a transcription tool is enough
At some point every tool gets pressure to grow sideways — features that look good in a comparison grid, a meeting recorder bolted on, an AI writing assistant nobody asked for. We have kept shrybe narrow on purpose. Paste a URL, get a transcript. Everything on our side is about making that specific thing faster and more accurate. Meeting recorders, audio editors, video-editing suites — those tools exist and they are well-made. We are not building the same things worse.
Three audiences benefit most from a tool this narrow:
- Podcasters drafting show notes, chapter markers, and quotable snippets from their own episodes.
- Journalists and researchers pulling timestamped quotes from public interviews, talks, and press conferences.
- Accessibility users — Deaf and hard-of-hearing viewers, plus people who simply consume text faster than audio.
A fourth group worth naming: developers and ML practitioners evaluating Whisper output before committing to a full pipeline of their own. The same architecture they would build with a few weeks of work, just hosted and free for casual use. We are aware of it and we don't mind. Even if you build something competitive after using shrybe, the transcripts you grabbed here helped someone, somewhere.
Contact
Email reaches us at hello@shrybe.com (catch-all, monitored). Public posts come through @shrybe_ai. If you found a bug, an edge case, or an embarrassing transcript artefact, please tell us — the bug report you send today is the fix the next person never has to encounter.
Drop in a URL, see what we mean.
Home, YouTube, podcast, or audio link. Same form, same 15-per-day cap, same absence of email field.