Cloud vs on-device, side by side
| Cloud dictation | On-device dictation | |
|---|---|---|
| Where speech is turned into text | A provider's servers | Your computer |
| Audio sent over the internet | Yes, every dictation | Never |
| Works without a connection | No | Yes, after setup |
| Account required | Usually | No |
| What decides speed and accuracy | Their servers and your connection | Your hardware |
| Typical pricing | Monthly subscription | Often one-time |
Why it got good enough
Open speech models such as OpenAI's Whisper made accurate transcription possible on ordinary computers. Its large-v3-turbo version runs comfortably on an NVIDIA graphics card or an Apple silicon Mac, and smaller versions run on a plain CPU. A small local language model can then tidy the transcript: punctuation, capitals, and dropping the ums.
What you need
- Windows: Windows 10 or 11, 64-bit, 16 GB of RAM recommended. An NVIDIA GPU lets you run the larger, more accurate model.
- Mac: Apple silicon (M1 or newer) on macOS 13 or later. 16 GB of memory runs the larger model.
- Disk space for the models, downloaded once during setup.
The honest trade-offs
- Your hardware sets the speed and accuracy ceiling. A recent machine does very well; an old laptop without a GPU uses a smaller model.
- Setup downloads a few models once, so the first run needs a decent connection.
- Clever cloud-only features, like syncing across devices, aren't available without a server.
How ScrinkleScribe does it
ScrinkleScribe runs Whisper and a local cleanup model on your computer and types the result wherever your cursor is, in any app. The only routine network request is a license check when the app starts, and it keeps working offline for 14 days. The privacy policy lists every request it makes.
Launching Nov 17
Hear on launch day: hello@scrinklescribe.com