On-device inference · offline by default

How it actually runs

From the moment you hit record to a titled, summarised, filed note sitting in a local database — every step happens on this device, and you can look up which engine tier each step used.

Download voNote

At a glance

What you said Next Wednesday at three, walk design through the new homepage, and swap out that icon we talked about.
Recording
Raw transcript

um… next Wednesday at three… walk design through the new homepage… and, uh, swap out that icon we talked about

The note

Homepage review with design

    Work Filed

    Transcript, title, key points and category written to the local database

    Front-end demo with a fixed sample sentence

    One recording, two processes

    After the recording, the real work is transcription and organising

    The two processes pick their engines independently

    They do not share one engine. Transcription may land on the fallback tier while organising still runs on the highest tier available — and the other way round. Which tier each step used is written into the note's processing receipt.

    The transcript is kept exactly as it came out, and organising only ever edits a copy — so a poor summary can never lose what you actually said.

    The five steps of one recording

    Scroll and the stage on the right follows along; click any step to jump there

    01 On device

    um… next Wednesday at three… walk design through the new homepage… and, uh, swap out that icon we talked about

    Homepage review with design

      · · · Work · · ·

      The transcript is kept exactly as it came out, and organising only ever edits a copy — so a poor summary can never lose what you actually said.

      Raw transcript
      The note
      Work

      Filed

      Three engine tiers

      There are three ways to do this, and the app always takes the highest one it can

      Two things we should say plainly about the lower tiers

      Tier 2 models can be downloaded and managed, but the inference backend is not wired into this version yet, so a download still resolves to another tier — once it lands, no setting changes and no re-downloading. Tier 3 is not a consolation prize: it runs on any device, and it is the layer that guarantees this pipeline always produces a result.

      Click a tier for the detail, or compare all three in the table below

      All three tiers, side by side

      One job, three ways of doing it, and every difference is in this table

      That last row is the point: tier 3 runs on any device, so getting a result is always guaranteed. The first two tiers decide how good the ceiling is.

      Which tier is this device on?

      The same code lands on different tiers on different hardware

      The model, the OS version and whether the language resources are in place all change the outcome. The check only reads this device and uploads nothing.

      Device and OS
      Bundled models
      Deduced result

      A web page cannot read your model or your OS version, so this deduces from the conditions you pick. For the real answer, run the one-tap check inside the app — that check also only reads your device and uploads nothing.

      In the app, open Settings → About voNote → How it works and tap “Check this device”. It tells you there and then which tier transcription and organising land on, and why a higher tier was not used.

      The network boundary

      Offline by default; going online takes your say-so

      Inside the device
      App network master switch
      Outside

      What the ledger records

      This switch is not a reminder — it is the actual interception point. Turn it off and anything needing the network fails there and then instead of quietly slipping through. Totals can be reset and every entry can be read one by one: don't take our word for it, go look.

      Common questions

      These seven come up most

      Put it on your phone

      Offline by default, never online without your permission