VOXA: a concept of a voice-first bank, with voice payments, money in chats and a shared dream jar

VITON¹³ builds concept sites to show its range in sectors where it has no client project to point to yet. VOXA is the one for banking and fintech: the site of a fictional bank of the future, in the genre of interactive feature stories where a preloader counts to ten and a camera flies through one 3D world, a feature per stop. The bank, its customers, balances and flights are invented, and the site is in English and Russian.

The brief we set ourselves: a bank you talk to and spend time in. A voice replaces the card number, friends replace the list of payees, and a trip, a dream or a split bill becomes something people do together, so the bank becomes a second place online after the messenger. And it had to be something a visitor can try, not only watch.

One page with ten features: voice payments, Voice ID, money in chats, flight-aware payments, shared dreams, bill splitting, a card lab with test and one-time cards, a friends’ feed with private amounts, an AI money coach and investment circles. Each feature has a diorama in the 3D world, a slide with a caption and an Explore panel that opens a working phone app at that feature; the menu lists all ten with thumbnails rendered from the scene.

The phone app runs on one shared state: money sent in a panel shows up in the chat, the balance and the test lab; friends reply and react; the dream jar fills while you watch; a bought ticket becomes a boarding pass with a live countdown, travel insurance and a local-currency note. A test lab gathers everything with commands to say or tap, and a microphone in the header steers the whole site by voice.

The closing scene cuts two 6-second films generated with Veo 3.1 (Alex asks VOXA to pay Anna for pizza; Anna laughs when it lands) together with the two phones on the page, then passes the microphone to the visitor. Every text is in English and Russian, and the films exist in both languages.

The world is one three.js scene: ten dioramas in a pastel studio whose colour flows from stop to stop, and a camera that flies between them along an arc with a slight lens breathing and follows the cursor. Every model is built in code, from phones whose screens are drawn on canvas to a white plane, a glass jar of coins, metal, glass and holographic cards and an arch carrying the members of an investment circle; the friends are illustrated portraits drawn in code, not photos.

Voice runs in the browser. Speech becomes text through the browser’s own recognition service and is parsed into an intent in English or Russian, Russian word forms such as «Ане» and «Максу» included. Voice ID records three phrases and keeps a voiceprint of the average spectral shape and pitch, computed in the tab and sent nowhere; spoken payments are checked against it, and a mismatch asks for Face ID. The site is built with esbuild into static files and served at /demos/voxa-bank/ with a header that allows the microphone on that path only.

What exists now is a working concept at tarasovvitalii.com/demos/voxa-bank/, built on 28 September 2026. A parser run the same day understood all 56 test commands, 28 in each language. We counted frames with requestAnimationFrame for three seconds, twice resting on a slide and twice during a camera flight, in headless Chrome with ANGLE Metal on a MacBook with an Apple M5 at 1600 × 1000 on a Retina scale: 60.1, 59.8, 60.0 and 58.1 fps. The page loads 0.24 MB of JavaScript with gzip, three.js included; the build is 4.3 MB, 3.3 MB of it the four film clips.

The same day Lighthouse 13.5 ran three times per device on the live address. The medians: performance 47 on mobile and 61 on desktop, accessibility 100, best practices 100; LCP 3.5 s on mobile and 0.5 s on desktop, CLS 0.005, and the page transfers 0.4 MB on load. What holds the mobile score back is the start of the 3D world: building ten dioramas and compiling their materials block a simulated phone for 15,754 ms. The machine was heavily loaded that day (load average around 50), so the live address should be measured again.

VOXA is a concept VITON¹³ built to show what we can make for banks and fintech products, not a commissioned project. The bank, the friends, the balances and the flights are fictional and no money moves; the two people in the films were generated with Veo 3.1 and Nano Banana Pro and are not customers. The site has no client, users, traffic or enquiries, and it proves nothing about conversion or retention.

Voice ID is a demonstration, not biometric security: it compares spectral averages, can confuse similar voices and was checked in headless Chrome with a synthetic microphone, not with real speakers. Speech-to-text needs Chrome, Edge or Safari. The frame rate comes from one machine on one day, and Lighthouse is a lab test that ran while the machine was under heavy load.

How to check a web studio’s portfolio →