Blog › Explainer
What on-device speech processing means for a planner app
2026-08-26 · Written by TrillionVolt, the maker of TimeLift
On-device speech processing means your voice is turned into text by the device (or its operating system) rather than being sent to the app maker's servers or an external AI API. For a voice planner this matters for three reasons: speech carries names, health, and appointments; skipping the server is what lets an app stay free; and no round-trip latency means a task appears the moment you stop talking. TimeLift uses the device's built-in recognizer and makes no external AI API calls.
Three tiers of speech processing
| Tier | Where the audio goes | Pros | Cons | Examples |
|---|
| ① Cloud AI | App server → external LLM (e.g., OpenAI) | Strong context understanding and command edits ("move dinner to next week") | Audio passes through third parties; usually paid or credit-metered; internet required | Day Flow, Voiset, SpeakToPlan, Motion |
| ② Built-in recognizer | The OS's speech recognition (Apple Speech on iOS, Google speech recognition on Android). No app server, no third-party AI | The app maker never holds your audio; no account; fast; can stay free | Understanding is rule-based inside the app, so command edits and reasoning are weaker. Depending on OS settings, the OS vendor's servers may be involved | TimeLift |
| ③ Fully offline | A model on the device only | Works without internet; strongest privacy | Accuracy and language coverage can trail the OS recognizer; larger app size | Some note-taking and transcription apps |
Exactly how TimeLift handles it
- Recognition — you talk on the call screen; the built-in recognizer in iOS or Android produces text. The audio doesn't go to TimeLift's servers, and no external AI API such as OpenAI is called.
- Splitting — "reply to emails, meeting at 3, should probably work out" is split inside the app, with rule-based logic, into three tasks with the filler trimmed.
- Time extraction — "at 3" is read and the task lands on the 3 PM block; untimed tasks are placed into your focus hours by Tami's auto-organize.
- The cap — the number of tasks scheduled today, N, is derived from your completion history; the overflow waits in a queue.
To be precise: the web version (timelift.app) uses the browser's speech recognition, which may route through the browser vendor's servers. On mobile, depending on OS settings, the OS vendor's (Apple's or Google's) recognition servers may take part. What TimeLift guarantees is "TimeLift and third-party AI services never receive your audio" — not "nothing ever leaves the device."
Why it matters
Privacy
Spoken to-dos contain things like "book the appointment at the clinic" or "pay back so-and-so." With cloud LLM processing, that audio and its transcript pass through at least two companies — the app maker and the AI provider.
Cost
A minute of cloud AI speech processing has a real price, which is why cloud-based planners are mostly $10–30/month subscriptions or come with limits like "30 voice minutes per month." The built-in recognizer has no such cost, which is what allows the core of an app to stay free.
Latency and trust
Live captions while you speak and a task appearing the instant you hang up are only possible without a round trip. The cloud approach's strength, in turn, is context-aware editing like "push that to next week." Which matters more depends on the user.
How to check which tier your app is
- Search the privacy policy for third-party AI, OpenAI, speech/voice data, transcript.
- See whether the voice feature requires signing in. Server-processed apps usually do.
- Try voice input in airplane mode. Built-in recognizers often work if the OS has downloaded its offline model.
- Look for an explicit sentence in the store listing, such as "uses only the device's built-in speech recognition."
Frequently asked questions
- What is on-device speech processing?
- Speech is turned into text by the device or its OS without being sent to the app maker's servers or an external AI API.
- Where does TimeLift process speech?
- The iOS and Android apps use the platform's built-in speech recognition; splitting and time extraction run inside the app. Audio never goes to TimeLift's servers or external AI APIs, and no account is required.
- So is TimeLift not "AI"?
- Utterance splitting, misrecognition correction, and priority placement run on rules and the app's own logic. It means no cloud LLM is used — not that nothing is automated.
More about TimeLift
Six voice planners compared