StoweA voice to-do list app lets you add tasks by speaking instead of typing. For most of their history these apps did exactly one thing with your voice: transcribed it. You said "call mum, gym tomorrow, book the dentist Monday", and you got that sentence back as text — one blob, in an inbox, waiting for you to split it into tasks yourself. In 2026, on-device AI changes the job description: the phone can now do the splitting, and it can do it without sending your voice anywhere.
Why was voice capture stuck at transcription?
Because understanding is harder than transcribing, and until recently the understanding had to happen on a server. Turning "gym tomorrow morning" into a recurring routine with a date attached is a structuring problem, not a dictation problem. Doing it well took a large language model, large language models lived in the cloud, and that meant every voice memo about your life made a round trip through someone else's computer.
So most apps stopped at dictation. Speech-to-text on the phone, structure left to you. The capture was fast; the sorting was still your job, and sorting is exactly the step people with a full head skip.
What does on-device AI actually change?
The structuring moves onto the phone. With Apple Intelligence and comparable on-device models, the same device that transcribes "call mum, gym tomorrow, post the passport forms" can also split it into three items, guess that one is a task, one is a routine, and attach "tomorrow" to the right one — with no network involved.
That collapses the old trade-off. You used to choose between a dumb-but- private voice inbox and a smart-but-cloud assistant. When the model runs locally, three things follow:
- Privacy stops being a trade-off. Nothing you say has to leave the device to be understood. There is no server-side transcript of your life to secure, subpoena, or leak.
- Offline stops being an edge case. A capture on a flight, in a basement gym, or on a hike works exactly like a capture at home. If the pipeline is fully on-device, airplane mode is not a degraded mode.
- Capture and structure become one step. The reason voice capture never quite stuck for many people is that it moved the mess from your head to an inbox, and the inbox still needed you. When the sorting is automatic, the brain dump arrives pre-sorted.
What should you look for in a voice to-do app in 2026?
Ask where the understanding happens, not just where the transcription happens. Concretely:
- Does it work in airplane mode — fully? Not "you can view your list offline" but capture, transcription, and structuring, all of it. This is the honest test of "on-device": turn the network off and try.
- Does speech become tasks, or a blob? Say three unrelated things in one breath. A 2026-grade app should give you three items with sensible types and dates, each editable, not one paragraph.
- Does it need an account? An app that understands you locally has less reason to require one. An account requirement usually means a server is involved somewhere.
- Does it handle how people actually talk? Half sentences, tangents, "oh and also". If capture only works when you dictate in clean commands, you will stop using it.
What is the honest state of the landscape?
Mixed, and moving. Dictation into a task inbox is everywhere — most major task managers take voice input through the system keyboard or Siri, and that path is fine if you do not mind sorting later. Cloud AI assistants that turn speech into schedules exist and are often capable, at the cost of sending your words to a server. Fully on-device speech-to-structured-plan pipelines are still rare, partly because capable local models only recently became something a developer can build on, and partly because most existing apps were architected around a backend years ago.
I sit in the last camp: Stowe, the app I build, transcribes and structures speech entirely on-device, works in airplane mode, and has no accounts and no server. I built it that way because I think that is where this category ends up — the phone is finally smart enough to keep your own thoughts on your own hardware, and once that is possible, sending them anywhere else needs a better justification than it used to have. But the shift is bigger than any one app: whichever tool you pick, in 2026 you no longer have to trade privacy for a to-do list that understands you.
If you want the specifics rather than the argument, there are honest side-by-sides of Stowe against two other apps that take tasks by voice: Stowe vs Voiset and Stowe vs Tiimo.
Frequently asked questions
Does voice capture work offline?
Only if both halves run on the device: transcription and structuring. Many apps transcribe locally but send the text to a server to be understood, so capture silently fails without a network. The honest test is airplane mode — if you can speak a capture and get structured tasks back with the radio off, it is genuinely on-device.
What is the difference between dictation and AI voice capture?
Dictation gives you back your words as text — one blob you still have to split into tasks yourself. AI voice capture does the splitting: "call mum, gym tomorrow, dentist Monday" becomes three separate items with types and dates attached. The difference is who does the sorting: you, or the model.
Can Siri turn speech into separate tasks?
Siri can add items to task apps by voice, one command at a time — "add milk to my shopping list" works well. What it does not do is take a rambling multi-item brain dump and split it into structured tasks; that takes an app built around a language model, on-device or otherwise.
Is voice capture private?
It depends entirely on where the audio and transcript are processed. If speech is understood on the phone, nothing needs to leave it. If the app sends audio or text to a server — most cloud AI assistants do — a transcript of your thoughts exists on hardware you do not control. Ask the vendor where the understanding happens; "we encrypt in transit" is not the same answer.