Technology & privacy
Offline vs cloud dictation, with no thumb on the scale
We sell an offline product, so here is the case for the other one stated properly. There is a real trade and pretending otherwise would waste your time.
Last updated
Why the trade exists at all
Speech recognition quality scales with model size, and model size is bounded by where the model runs. A data centre can run something far larger than a laptop can. That is the entire root of the difference, and everything else follows from it.
What has changed in the last few years is how much that gap matters. Local models used to be noticeably worse on ordinary speech. On Apple Silicon, with current models, they are not — on clean audio from a decent microphone the difference is small enough that your vocabulary matters more than the model.
Where the gap is still real: noisy environments, strong accents in a second language, rapid code-switching between languages, and unusual technical vocabulary. If your audio is difficult, cloud still leads and will for the foreseeable future.
The full comparison
| Cloud | Offline / on-device | |
|---|---|---|
| Peak accuracy on hard audio | Better | Close on clean audio, behind on hard |
| Language coverage | Very broad | Narrower |
| Audio transmitted | Yes | No |
| Works with no connection | No | Yes, after the model download |
| Latency | Network round trip plus processing — usually fast, occasionally not | Processing only — the same every time |
| Behaviour on bad Wi-Fi | Degrades or stops mid-sentence | Unaffected |
| Cost model | Subscription, because inference costs the vendor money | One-off or subscription; the compute is yours |
| Battery | Minimal local cost | Uses your own silicon |
| Works on every device | Yes, if the vendor ships there | Only where the app runs |
| Sync between devices | Usually | Nothing to sync |
| Still works if the vendor disappears | No | Yes |
| Third party to assess for compliance | Yes | None for recognition |
The case for cloud, made properly
- It is genuinely more accurate where accuracy is hard. A café, a car, a second language, a strong regional accent. Dismissing this would be dishonest.
- Language coverage is not close. If you dictate in a language a small local app does not ship, there is no argument to have.
- It works on your phone. Most people's dictation happens on more than one device, and local apps are per platform.
- No model download, no disk space, no cold start. Install and go.
- It improves without you doing anything. The vendor upgrades the model; you notice one Tuesday.
The case for offline
- The audio is not transmitted. Architecture rather than policy, and demonstrable in two minutes. For confidential work this is not a preference, it is the requirement.
- It works where there is no signal. Planes, trains, basements, client sites, hotels.
- Predictable latency. As above.
- You can buy it outright. Only possible because the vendor is not paying for every dictation.
- It keeps working if the company stops. The model and the app are already on your disk.
- Nothing to assess for compliance. No processor, no data transfer, no agreement to negotiate.
How to decide
- Is there anything you dictate that must not be transmitted? If yes, that decides it. No accuracy advantage compensates for a requirement you cannot meet.
- Do you regularly work without a connection? Flights, trains, site visits. If yes, cloud dictation is not a tool you can rely on.
- Do you dictate on a phone as well as a computer? If yes, and it matters, cloud is the practical answer today.
- Is your audio difficult? Noisy room, second language, strong accent. If yes, test both before committing — this is where the gap is real.
- Do you object to renting it indefinitely? Only offline products can reasonably offer a one-off licence.
The hybrid nobody mentions
You can run both. Give them different shortcuts. A common arrangement: a local app for anything confidential and for travel, a cloud app on the phone or for difficult recording conditions. Nothing about either architecture prevents it, and free tiers make it cheap to try.
Questions
Is offline dictation less accurate?
On clean audio from a decent microphone in a reasonable room, the gap with large cloud models is small enough that custom vocabulary matters more. In noisy conditions or with a strong accent in a second language, cloud models retain a real advantage.
Is cloud dictation faster?
Often lower latency on a good connection, and much worse on a bad one. Local latency is usually slightly higher and identical every time. For an input method the consistency is generally worth more than the average.
Does offline dictation use more battery?
Yes — the inference runs on your own silicon rather than a server's. On Apple Silicon it is modest, but it is not zero, and it is noticeable across a long session with no power.
Can I use both?
Yes. Different shortcuts, no conflict. Local for confidential and offline work, cloud for the phone or for difficult audio.