SayLocal›Guide: local vs cloud
Local vs cloud speech recognition: the differences
Every dictation tool has to turn a recording into text. What matters is where that happens: on a server or on your own computer. That affects privacy, offline use, speed and hardware.
For Mac and Windows · Trial with an email address, no credit card
The difference in one sentence
With cloud speech recognition the recording is sent to a server and turned into text there. With local speech recognition this step runs on the computer you are speaking into – the recording never leaves the device.
Comparison
| Cloud | Local | |
|---|---|---|
| Recognition | On the server | On your computer |
| Sent over the network | Every recording | No recording |
| Offline | No | Yes |
| GDPR processing agreement | Usually needed | Not needed for the recognition |
| Hardware | Hardly any requirements | A GPU helps a lot |
| Download | Small | Speech model once (SayLocal: about 550 MB) |
How fast is local recognition?
It depends on the hardware. With SayLocal on a Windows PC with a Vulkan-capable graphics card, text appears after about 0.2 seconds; on the processor alone it takes a few seconds. On Macs with Apple silicon recognition runs on the chip’s graphics unit. Intel Macs are noticeably slower locally.
That is why SayLocal offers an optional cloud recognition on Intel Macs and PCs without a suitable graphics card – off by default, only with an explicit tick. On Macs with Apple silicon and PCs with a suitable graphics card this option does not exist.
And accuracy?
Accuracy depends mainly on the speech model, not on whether it runs locally or on a server. For both, the hard part is usually names and technical terms. SayLocal compensates with context: it reads names from the window you are replying to (in working memory only) and remembers words you correct.
Measured on 16 Oct 2026: 14 German test sentences with 104 words, including names, places and technical terms, MacBook Pro with M3 Pro. The same speech recognition without SayLocal: 5.8% word errors; with SayLocal: 0%. Results vary with other voices, microphones and texts.
What to check when an app says “local”
- Does the recognition really run locally – or only the app, which then uploads recordings?
- Is there a cloud option, and is it off by default?
- What still goes over the network – license check, updates, usage figures – and which of it can be switched off?
- Is screen content used, and if so: stored, or only kept in working memory?
- Does the app require an account?
SayLocal’s answers are on private dictation app.
Frequently asked questions
Is local speech recognition GDPR compliant?
Local processing avoids transmitting recordings, so no data processing agreement is needed for the recognition itself. Whether a solution meets your obligations overall is for you or your data protection officer to assess – this is not legal advice.
Is local speech recognition slower?
Not noticeably with a GPU: SayLocal delivers text after about 0.2 seconds on a PC with a Vulkan-capable graphics card. Only on the processor or on Intel Macs does it take a few seconds.
Does local speech recognition work offline?
Yes, that is one of its advantages. SayLocal only needs a brief connection once or twice a day for the license check and keeps working for up to 7 days without one.
How much storage does a local speech model need?
For SayLocal a one-time download of about 550 MB, about 1 GB of free storage and 8 GB of memory; while dictating the app uses about 800 MB on the Mac.
More about SayLocal
Updated 23 October 2026 · Conto IT, Itzehoe
