Blog

The app can now show a banner from the site — a release note, an image, a link — that you can open or dismiss. And there's a switch in the drawer: share anonymous usage, off by default. If you turn it on, the app sends counts only — never what you said, never audio.

That switch exists because the promise is real: nothing leaves your device unless you turn something on. The website you're reading runs on the same small server the app already trusted.

The charts were hairline bars with no labels — pretty, useless. Now they're chunky bars with real axes: hours on the day view, weekdays on the week view, dates on the month. And a monthly estimate that projects what your pace would cost over thirty days.

Small thing, big difference: the day view used to quietly show a month of data under a "today" label. It says "day cost" now, and it means it.

The usage screen used to estimate what cloud transcription cost — the math was right but the number wasn't real. The provider already tells us exactly what each segment costs, down to the smallest fraction of a cent.

So now that's what you see: the real bill, not our reconstruction of it. Old estimates stay where they are, clearly marked, but everything new is the genuine ledger.

The usage page started as counters. It became a mirror: your most-used words as a cloud, how often you reach for "um" and "basically", whether you ask questions or make assertions, how your utterances are sized.

It also watches your hedges — "I think", "maybe", "does that make sense" — forty of them now, because the first list missed how people actually soften what they say. We named the metric "softening" because that's what it is.

Early on, every screen kept its own numbers — and they never quite agreed. So we made one place where every dictated segment lands: words, duration, which engine served it, what it cost. Everything the usage screen shows is read back out of that single record.

It lives on your phone and nowhere else. The day we wiped it in testing, the whole screen went politely blank — which is exactly what wiping should do.

Touching the mic and inserting text are different gestures, so they got different haptics — different strengths, different feels. A light tick for a tap, a firmer click when words land. Insert defaults to gentle: a confirm, not a celebration.

Then the crash. Haptics worked everywhere except where it mattered — because we'd never actually asked the phone's permission to vibrate. One missing line, one humbling fix. The boring parts of a feature are still the feature.

There's no app store between you and the next build. The app quietly checks its own home on the web, and when there's something newer, it downloads it and offers the install sheet itself. If it can't reach home, it says so instead of pretending everything's fine.

Even that had a learning curve: at one point our own host was refusing the app's polite knock until it learned to introduce itself like a browser. Now the knock works, and a cancelled install doesn't nag you twice.

Some people want the bigger engines. Fine — but we didn't want to be a reseller metering your speech. So the app takes your own key to the transcription marketplace, and your audio goes straight from your phone to them under your account.

We never see it, never touch it, never mark it up. The bill is yours at cost. That's the whole arrangement.

The project started under a working name, and it got a real one mid-build. Renaming sounds cosmetic until you do it: the app, the file names, the update server, the domain — everything moves at once, and anything that slips keeps silently asking the old name for updates.

So the app learned to check both doors during the handover — ask the new home first, fall back to the old one if it's still answering. Nobody's install got orphaned in the move.

Every local model downloads straight from public model hubs — we host nothing, so distribution costs us nothing, so the app can be free for real. Not free-with-ads, not free-for-a-trial.

When we added models that only existed on one hub, we wired in their real repositories rather than pretending. If the source moves, the app follows the source — not the other way around.

The picker grew filter chips the way every catalog does — one per idea, because more filters feel like more power. Then we tapped through them: half of them narrowed the list to the same three things. They weren't filters, they were decoration.

So the tags got sorted into groups that actually separate the models — language, speed, special skills. One tap, one filter, and every chip tells you how many matches you'll get before you press it.

People's phones and languages are different, so instead of one engine there's a catalog. Tiny and instant, or heavy and precise — five on-device choices to start, more behind the curtain, plus cloud engines you can plug your own key into.

Each card tells you its size, its languages, its temperament. Pick once or switch anytime — it remembers you either way.

The first real decision: no app to open, no screen to manage. A small floating pill that appears when a text field has focus — tap, speak, text lands where you're typing, gone. One gesture end to end.

Everything about it had to say "always there, never in the way". Color and motion carry the state — red means it's listening, shimmer means it's thinking. Words were deliberately left out.

Voice typing should be the fastest way to put words in a text field. But every option asked for a bad trade: the good one wanted your audio and your screenshots in their cloud, for a monthly fee. The free ones were degraded versions that kept getting worse.

Meanwhile your phone got fast enough to do the whole thing itself. So the question became: what if the best voice keyboard was the one that never needed to send your voice anywhere?