A speech translator and phrasebook for Kazakh, Russian, and English that keeps working when the signal doesn't — built for travel across Central Asia, and doubling as a way to actually learn the phrases.
Translation apps assume a connection, which is exactly what you don't have on the road in Central Asia. The brief was a fast Kazakh–Russian–English translator that works offline first, is honest about how confident it is in each result, and helps you learn the language rather than only relay it.
A React + TypeScript app pairs the browser's speech recognition and synthesis (Web Speech API, with a dedicated voice per language) with a local translation engine. A curated phrase pack and lexicon — grouped by travel, transport, daily, and medical situations — resolve common phrases instantly and offline. Every result carries a confidence score and a source label, so you know whether you're getting a verified phrase or a best-effort lexicon match.
The same phrase pack powers a learning mode: phrases are organised by category for study, each with native script and audio in all three languages. The interface is built around a quick speak-and-swap between Kazakh, Russian, and English, so a conversation can bounce back and forth without digging through menus.
Soile stays useful in exactly the moment most apps fail — no signal, mid-conversation — and its honesty about confidence means you can trust a result or fall back to a verified phrase. As a bonus, the phrasebook foundation makes it a low-pressure way to pick up Kazakh.
A live build and a short write-up of the offline engine and confidence model can be shared on request.