"AI dictation" usually means one thing under the hood: your voice goes to a vendor's servers, their models do the work, text comes back. The AI is real — the trade-off is that your words make a round trip through someone else's infrastructure, forever, on a subscription. An Apple Silicon Mac doesn't need the round trip. Tota runs the entire AI pipeline — speech recognition, a formatting language model, and speaker identification — on the machine you already own.
Three Models, Zero Cloud
Tota's AI stack is three on-device models doing three jobs:
- Speech recognition — OpenAI's open-source Whisper models, run through WhisperKit, Argmax's framework built specifically for Apple Silicon. This is the same class of model behind most cloud dictation services — running on your Neural Engine instead of their servers.
- Smart Formatting — an optional cleanup pass by Qwen 3, an open-source language model compact enough to run locally via Apple's MLX. It fixes typos, misheard homophones, punctuation, capitalisation, and stutters, and formats spoken lists properly. One download (~2.3 GB), then it's yours.
- Speaker identification — when you transcribe audio files, diarization runs on-device too, labelling who said what.
Corrects, Never Rewrites
The fear with AI-in-the-loop dictation is the AI getting creative. Tota's formatting model is deliberately constrained: it fixes errors — it never paraphrases, summarises, changes numbers, or acts on instructions it hears in your dictation. Every formatted result must pass safety checks (length, overlap with what you actually said, processing time) or Tota discards it and pastes the raw transcription instead. You also direct layout by voice — say "new paragraph", "new line", or "new list" and the words themselves disappear from the output.
AI That Adapts to You
Cloud AI dictation is one model for everyone. Local AI can be personal without being invasive: Tota's auto-learning dictionary picks up the corrections you make, and custom glossaries bias the speech model toward your project names, jargon, and clients' names — switching automatically per app. None of that profile ever leaves your Mac, because there's nowhere for it to go.
Local vs Cloud AI Dictation
| Local AI (Tota) | Cloud AI (Wispr Flow, Aqua, Willow) | |
|---|---|---|
| Where your voice is processed | On your Mac | Vendor servers |
| Works offline | Always | No |
| Latency | Your hardware, no network hop | Network-dependent |
| Why it's priced that way | £19.99 once — your Mac does the compute | Subscriptions — every word costs the vendor server time |
| Account required | No | Yes |
| Data available to train on | None — nothing leaves the device | Varies by vendor and settings |
The economics are worth dwelling on: cloud vendors charge monthly because every word you dictate costs them GPU time. When the models run on your Mac, the marginal cost of your dictation is zero — which is why genuinely local apps can charge once. The full pricing comparison across eight apps is in best dictation apps for Mac.
What You Need
Any Apple Silicon Mac (M1 or later) on macOS 14+. No GPU rig, no API key, no account. Models download once — after that, flight mode is a supported configuration. How fully offline operation works, and which competing apps can and can't claim it, is covered on the offline dictation for Mac page.
Frequently asked questions
What does the AI in Tota actually do?
Three jobs, all on your Mac: Whisper-class speech recognition turns your voice into text, an optional local language model (Smart Formatting) fixes typos, punctuation, and lists, and a diarization model labels speakers in transcribed audio files.
Do I need an internet connection or a special GPU?
No internet after the one-time model downloads, and no extra hardware — Tota uses the Neural Engine and GPU already in every Apple Silicon Mac, via Apple's MLX and WhisperKit frameworks.
Will the AI rewrite or summarise what I said?
No. Smart Formatting corrects, never rewrites: it fixes errors and formatting but doesn't paraphrase, change numbers, or act on instructions in your dictation. If its output fails Tota's safety checks, the raw transcription is pasted instead.
Is local AI dictation as accurate as cloud AI dictation?
For English and major languages on Apple Silicon, local Whisper-class models are competitive with cloud services. Tota also adapts to you — custom glossaries and an auto-learning dictionary — which is where one-size-fits-all cloud models tend to miss.
Which AI models does Tota run?
OpenAI's open-source Whisper models via WhisperKit for speech recognition, and Qwen 3 — an open-source language model running through Apple's MLX — for Smart Formatting. Both run entirely on-device.
