Say It or Snap It — Either Way, You're Not Typing

The short version

People don't abandon finance apps because the analytics aren't smart enough — they abandon them because logging a transaction takes too many taps, in a workflow designed around English and around typing. We built two ways around it: say it out loud — "coffee, six fifty, cash" — in English, Russian, German, Spanish, or French, or just photograph the receipt. Either way, ExtMoney drafts the transaction and you confirm it; nothing gets written to your ledger on its own.

Typing is the tax nobody budgets for

Every way of tracking money — envelope, zero-based, simple categorization, a running net-worth figure — eventually needs the same raw material: an accurate, timely record of what happened. The method isn't usually the problem. The tax you pay to feed it is: opening the app, finding the right account, typing the amount, picking a category, maybe adding a note, confirming the date.

Missing that tax for a week doesn't mean losing a week of data. It means facing a backlog that feels like an accusation, which is exactly the moment most people quietly stop opening the app at all. Voice and photo entry are our answer to that specific moment — not a novelty feature, a direct attack on the actual failure mode.

Why five languages, and why that's harder than it sounds

We support English, Russian, German, Spanish, and French as first-class languages for voice, not as an English product with translated buttons. That distinction shows up in places that are easy to miss until they break.

Early in testing, our system for filtering out background noise and non-transaction speech — telling "add fifteen dollars for coffee" apart from "remind me to call my mother" — had been tuned against English phrasing and was quietly rejecting perfectly clear Russian and French transaction requests as noise. Multilingual isn't a translation layer sitting on top of an English core; every piece of logic that touches the raw words a person says has to be built, and tested, against all five languages from the start, or it fails silently in exactly the languages it was supposedly built for.

Or just take a photo

Not every purchase lends itself to narrating out loud — a stack of receipts from a trip, a paper bill from a landlord who doesn't do digital, a purchase made somewhere too loud or too public to talk to your phone. For those, point the camera at the receipt. The amount, date, and merchant get read automatically, and it lands in the same draft-then-confirm flow as a voice entry.

The two methods aren't really different features so much as two entry points into the same pipeline: something turns messy real-world input — sound or a photograph — into a structured draft, and nothing gets written to your books until you've looked at it.

The part that has to be a human, not a model

We route audio through AssemblyAI as our primary speech-to-text provider, chosen for strong accuracy across all five languages and an EU-hosted processing option, with a small number of cheaper or faster alternatives as fallback if the primary is slow or unavailable. That's an availability decision, not an accuracy one — a slow or down primary provider never means "no voice input today," it means falling back smoothly to another option.

But no matter how good the underlying transcription or image recognition is, we don't let a model decide your transaction and quietly save it — for voice or for a receipt photo, the rule is the same. Every parsed transaction goes through deterministic, code-level checks before it touches your ledger: "yesterday" gets resolved to an actual date by code, not by however the model felt like interpreting it that day; amounts get validated as real numbers in a sane range; a merchant, account, or category has to match something that already exists in your data, or the app flags it for you to confirm rather than quietly inventing a new one. The model's job is turning speech or a photo into a draft. Deciding whether that draft is trustworthy enough to save without asking is a job we kept out of its hands entirely.

Fast when it's confident, careful when it isn't

If what you said or photographed was clear and unambiguous — amount, category, and account all matched cleanly — it saves immediately. If anything was ambiguous — an unclear amount, a faded receipt, a category that doesn't quite match your existing tags — it surfaces for a quick confirmation instead of guessing. That's a deliberate trade of a little speed for a lot less risk of a wrong number sitting silently in your financial history, which matters more in a finance app than in almost any other category this kind of input gets used for.

Practical recommendations (for evaluating voice or photo entry in any finance app)

• Ask what happens when the app isn't sure. Does it silently guess, or does it ask? That answer tells you more about how trustworthy the feature actually is than any accuracy percentage in the marketing copy.
• Test voice in whatever language you'd actually use it in, not just English, even if the app claims multilingual support. Translated interfaces and genuinely multilingual logic are not the same thing.
• Test the receipt scanner on something messy — a crumpled receipt, faded thermal paper, a receipt in a currency other than your default. That's where the real gap between demo and daily use shows up.
• Ask what happens to the audio or photo itself after it's processed. Voice recordings and receipts are sensitive by nature; whether they're retained, and where, is a fair question to ask any app that listens to or looks at your purchases.

Where we'd push back on ourselves

• Both channels will misfire, at some rate above zero. Background noise, accents, and code-switching for voice; bad lighting, faded ink, and unfamiliar layouts for receipts — real failure modes we design around rather than pretend away, which is the entire reason the confirmation step exists.
• The confirmation step is a real trade-off, not a free lunch. It means neither channel is fully hands-off; some entries still need a tap. We think that's the right trade for financial accuracy, but it is a trade.
• Running multiple speech-to-text providers is more moving parts than one. It buys availability and cost flexibility, but it's genuinely more to build, test, and keep working correctly than a single-provider integration would be.
• We're just launching, so real-world accuracy across five languages, accents, receipt formats, and phone cameras at real scale is still ahead of us, not behind us.

---

If "say it once, in your own language, or just snap the receipt, and only get asked when it actually matters" sounds like the finance app you've been waiting for, start for free at ExtMoney — we built this one out loud, and in focus, too.