Local AI vs cloud AI for cleaning up dictation
When a dictation app uses AI to turn your raw transcript into clean text, that model can run two ways: locally on your Mac or in a company's cloud. Local keeps your words on your machine, works offline, and costs nothing per use. Cloud models can be a little more capable on complex rewrites, at the cost of sending your text to a third party. Relay supports both, so the choice is yours.
Last updated . Prices and vendor claims come from the sources listed at the end of the page; re-check before purchase.
Transcription and cleanup are two separate steps. Relay transcribes your voice on-device with Whisper either way. The question here is the second step: when you ask AI to fix punctuation, remove filler, or rewrite a rambling thought into a clean sentence, where does that happen?
The case for local AI
- Privacy. Your dictated text often contains the sensitive stuff: client emails, code, medical or legal notes, half-formed ideas. A local model processes all of it on your Mac, so none of it is sent to anyone.
- No per-use cost. Local models are free to run once installed. Cloud APIs charge per token, which adds up if you dictate all day.
- Works offline. On a plane or bad Wi-Fi, a local model still cleans up your text.
- No lock-in or training concerns. Your text isn't logged on a vendor's servers or potentially used to train future models.
The case for cloud AI (closed models)
- Capability. Frontier closed models (the latest Claude or GPT, for example) are still ahead of small local models on hard rewrites, tone changes, and reasoning-heavy edits.
- No local hardware needed. Cloud runs on someone else's machine, so a lighter Mac still gets top-tier output.
The honest trade-off
| Relay | Cloud / closed AI | |
|---|---|---|
| Your text stays private | Yes (local) | No, sent to the provider |
| Cost per use | None | Per-token API charges |
| Works offline | Yes | No |
| Peak capability on hard rewrites | Good (small local models) | Best (frontier models) |
| Needs a capable Mac | Yes (Apple Silicon) | No |
How Relay handles it
Relay's transcription is always on-device. For the optional AI cleanup, you choose: run a local model for full privacy and zero cost, or plug in your own API key if you want a frontier cloud model for the heaviest rewrites. Either way you're in control, and nothing is sent anywhere you didn't choose. For most everyday dictation, a local model is more than enough, and your words never leave your Mac. See also using your own API key and is voice dictation private.
Frequently asked questions
Are local AI models good enough to clean up dictation?
For everyday tasks (fixing punctuation, removing filler, light rewrites of emails and messages) yes, small local models do this well. For the most demanding rewrites or reasoning, frontier cloud models still have an edge. Relay lets you use either.
Why use a local LLM instead of a cloud one?
Privacy, cost, and offline use. A local model processes your text entirely on your Mac, so sensitive content never leaves your machine, there's no per-token cost, and it works without internet. Closed cloud models can be more capable but require sending your text to a third party.
Does Relay send my text to OpenAI or Anthropic?
Only if you choose to. Relay transcribes on-device and can do AI cleanup with a local model (nothing sent anywhere). If you'd rather use a frontier cloud model, you can plug in your own API key, in which case the text you choose to clean up goes to that provider under your own account.
Is local AI cheaper than cloud AI?
Yes, over time. Local models are free to run once installed, while cloud APIs charge per token. For heavy daily dictation that difference adds up, which is why local is the cost-effective default for routine cleanup.
Dictate with AI that stays on your Mac
Relay: on-device transcription, with local or bring-your-own-key AI cleanup. Free during beta.
Download for MacFree during the beta · Apple Silicon (M1 or later) · Your voice never leaves your Mac
Get a license key and release notes by email (optional; the download does not require it)