Built for the file you already have somewhere
Most transcription tools want you to make an account and wait while a meter ticks. shrybe does it differently: if your audio already lives at a URL – a podcast CDN, a static server, a Dropbox file link – paste that link and we fetch the file from there. If it only lives on your device or in Google Drive / iCloud, hit Upload (up to 100 MB). Either way there is no account on our side to associate it with.
Three steps
- Get the file or its link. A direct .mp3 / .m4a / .wav URL from a podcast host, CDN, or your own server pastes as-is. File only on your device or in a cloud drive? Use the Upload button – up to 100 MB.
- Paste or upload it. Linked files are pulled straight from the source URL; uploads travel over an encrypted connection and count the same as a link – one of your 15 daily transcripts.
- Read the transcript. Whisper writes timestamped paragraphs. Copy a passage, download the whole thing as .txt, or scroll through it in the browser. Typical 20-minute recording finishes in under a minute end-to-end.
The formats Whisper actually handles
Whisper does not care what container your file arrives in, and the reason is written into the paper: before anything reaches the model, "audio is re-sampled to 16,000 Hz, and an 80-channel log-magnitude Mel spectrogram representation is computed" (Radford et al., 2022). A 320 kbps MP3, a studio WAV and a phone voice memo all become the same 16 kHz mono representation. That is why a smaller file does not mean a worse transcript, and why re-encoding your audio to "help" achieves nothing. Anything our backend can decode works:
- MP3 / M4A / AAC – the everyday formats from voice recorders, podcasts, and phones.
- WAV / FLAC – uncompressed and lossless. Bigger files, identical accuracy to compressed.
- OGG / Opus / WebM – common from Discord, WebRTC, and modern web recorders.
- 3GP / AMR – older phone-recording formats. Still supported, accuracy mostly limited by the recording quality, not the codec.
What actually decides how long it takes
Not the format, and not the file size. Whisper reads audio in fixed windows – the paper trains on 30-second segments and handles long recordings by "consecutively transcribing 30-second" slices (Radford et al., 2022) – so the work scales with how long the recording is, never with its bitrate. A 60-minute lossless WAV and a 60-minute MP3 are the same amount of work. Compressing a file before you send it saves you upload time and nothing else.
The other half is the checkpoint. The turbo model here cuts the decoder from 32 layers to 4, landing at 809 M parameters. That is where the speed comes from: a shallower decoder writing down what the same encoder heard. It is the trade every fast Whisper deployment makes, shrybe included.
What this is good and bad at
Good: single-speaker recordings – lectures, voice memos, a recorded interview where one person at a time speaks. Clean English sits at the good end of Whisper's published range: across the fourteen English datasets in Radford et al., 2022, Table 2, the Large V2 model runs from 2.7% word error on clean read speech to 36.4% on far-field meeting audio, averaging 12.8%. What moves that number is how the audio was captured, not how fast anyone talks. Across other languages the picture is uneven by OpenAI's own account – the model card states plainly that the models "perform unevenly across languages", with lower accuracy on low-resource ones.
Less good: multi-party meetings with overlapping speech. Whisper doesn't perform speaker separation (diarisation) by default – two people talking over each other become one mangled paragraph. Loud background music, heavy echo, or recordings made across a large room with one phone in the middle also lose words. If your recording sounds bad to you, the transcript will reflect that.
The mitigation: if you control the recording, record close to the speaker, in a room with soft surfaces, and ask people not to talk over each other. None of those tips require fancy gear; they make more difference than any model upgrade would.
Privacy and retention
shrybe fetches the audio from your cloud share, runs it through Whisper, and returns the transcript to your browser. We cache the audio for at most 24 hours to allow retries if the connection drops; after that a nightly job deletes it. The transcript text itself is never stored on our side beyond delivery. The free tier rate-limits by IP at 15 jobs per 24-hour window – that is the only telemetry we keep, and it rotates out after 30 days.
You can revoke the share link in your cloud drive the moment the transcript appears. Most users do.
Frequently asked
Which audio file formats does shrybe support?
Every common audio container: MP3, M4A, AAC, WAV, FLAC, OGG, Opus, WebM-audio, and a few rarer ones. All of it is decoded and re-sampled to 16 kHz mono before the model sees a thing, so the container almost never matters. A decent rule of thumb: if a desktop player like VLC opens your file, shrybe can transcribe it.
Can I upload a file directly from my computer?
Yes – free. Use the Upload button next to the URL field: drag-and-drop or browse, up to 100 MB per file. An upload counts as one of your 15 daily transcripts, same as a link.
Can I paste a Google Drive, Dropbox, or iCloud link?
Dropbox links that point at the file itself work – we fetch what they serve. Google Drive and iCloud share links open a viewer page rather than the file, so our fetcher cannot use them; for files stored there the Upload button is the faster path anyway. Any URL ending in the audio extension (.mp3, .m4a, .wav…) works from any host.
Is there a maximum file size or duration?
The free tier caps each transcription job at roughly 75 minutes of audio. There is no hard MB limit because we re-encode the audio to a compact format before Whisper runs, but very large files (over 1 GB) will time out during the fetch step before transcription even starts.
How accurate is Whisper on lectures, interviews, or recorded calls?
It depends on the recording far more than on the subject. Across the fourteen English datasets in Radford et al. (2022, Table 2), Whisper Large V2 runs from 2.7% word error on clean read speech to 36.4% on far-field meeting audio, averaging 12.8%. A lecture recorded close to the speaker sits near the good end; a phone call with echo and several people talking sits near the other one. Cross-talk is the consistent failure mode – Whisper does not separate voices, so overlapping speech becomes one mangled paragraph.
What happens to my audio after the transcript finishes?
The cached file is deleted within 24 hours by a nightly job. The transcript text is delivered to your browser and not retained on our side. We never train on free-tier traffic.
Will shrybe work with audio from Voice Memos or a phone recording?
Yes, once you upload the .m4a (iPhone Voice Memos default) or .3gpp (Android) file to a cloud host with a public share link. Phone recordings often need a louder normalisation pass before transcription – Whisper handles soft audio reasonably well, but very quiet recordings of meetings across a table will lose words.
Can I transcribe a Twitter Spaces or audio space recording?
If you can grab the audio file URL from your Twitter/X downloads, yes. Twitter does not expose Spaces audio cleanly through their public links anymore; you may need the host to share an exported recording with you.
Sources
The technical claims above are quoted from primary sources, not from secondary write-ups:
- Radford, Kim, Xu, Brockman, McLeavey, Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision (2022, arXiv:2212.04356) – the 16,000 Hz re-sampling and 80-channel Mel spectrogram, the 30-second segments, and the English word-error-rate spread in Table 2.
- openai/whisper-large-v3-turbo model card – the decoder cut from 32 layers to 4, the 809 M parameter count, and the "perform unevenly across languages" caveat.
Ready when the file is.
Paste your audio URL above. If you'd rather transcribe a YouTube video or a podcast episode, those have their own pages.