WMSOFT Lab is building an AI-powered language learning game where every scene is a quest. You speak, an AI character answers in character, and the next line is scored and adapted to you — in real time, out loud.
Interface mockup of a conversation quest. Layout and scoring shown for illustration; screens may change before release.
Most language apps test what you can recognise. This one keeps you in conversation until the scene is complete — then shows you exactly what to fix and brings it back when you are about to forget it.
Every lesson is a mission inside a living scene. An AI character has its own goal, mood and vocabulary level, and it answers what you actually said — not a fixed script. Ask for the wrong dish and the waiter pushes back; improvise and the story follows you. Each quest has objectives you can win or lose, from ordering breakfast in Lisbon to a job interview in Seoul.
Speech recognition scores your turn within seconds on pronunciation, stress, rhythm and word choice. Feedback is specific rather than a percentage: which syllable landed wrong, which phrase was understandable but unnatural, and a native-sounding alternative you can retry immediately. A per-phoneme profile tracks what is genuinely hard for speakers of your first language.
Levels, streaks, collectible characters and scenario unlocks carry the habit past the first week. Underneath, a spaced-repetition scheduler resurfaces each word and phrase at the point you are about to forget it — built from the mistakes you made in real conversations, not from a generic word list that ignores what you already know.
One account and one learning profile across iOS, Android and web. A cloud-native backend streams audio through a model pipeline for transcription, dialogue and speech synthesis, with a latency budget set per conversational turn. Model routing is provider-agnostic, so a speech or language component can be replaced without touching the client apps.
Choose a scene, a level and a target language. The scenario states its objectives before you start talking.
Hold to speak. The AI character replies in voice, in character, and reacts to what you actually said.
Pronunciation, grammar and word choice are marked per turn, with a native-sounding alternative to retry.
XP and unlocks move you forward; a review queue keeps your own mistakes coming back at the right moment.
Talking to a character only feels real if the reply lands fast. The architecture is organised around that constraint: stream everything, keep per-turn context, and isolate the parts that will change as models improve.
One codebase for the interface, native audio capture and playback on each platform.
Authenticated streaming sessions with partial transcripts and turn-level timeouts.
Speech-to-text, a scenario-aware dialogue model, and speech synthesis, plus a separate pronunciation scorer.
Versioned quest content, per-learner progress, review scheduling and privacy-scoped audio handling.
Audio is transcribed as it arrives instead of after the learner stops speaking, so the character can begin answering before the turn is fully decoded. The target is conversational, not batched.
Each quest keeps objectives, character state and the learner's level in structured context, so the character stays consistent and can nudge the learner back on task.
Speech and language models are accessed through an internal interface. Components can be swapped, benchmarked per language, or run in different regions without changing the client.
Quests are authored in a structured format and versioned per language and level, so adding a language is a content task rather than a code change.
Voice data exists to score a turn. Retention is a setting, not an afterthought, and account deletion is designed to remove the learner profile rather than archive it.
This is a company at the start, and we describe it that way. The conversation engine and quest framework are in active development; the content library, the apps and the store releases come next.
Conversation engine and quest framework in active development. Content pipeline for the first three languages. Closed beta preparation for iOS and Android.
Invited testers on both mobile platforms. Pronunciation scoring calibrated against recorded learner speech. First set of quests across travel, daily life and work.
App Store and Google Play release, a web companion for practice and review, and subscription tiers for individual learners.
Classroom and institutional pilots, additional languages, and an authoring tool for teams that want to build their own scenario content.
Status as of November 2026. Dates on this roadmap are indicative and may move as testing progresses; nothing here should be read as a committed release date.
WMSOFT Lab is the product lab of Waymaker, an independent software studio based in the Republic of Korea. The lab exists to design, build and test products that would have been impractical before modern AI — and to put them in front of real users quickly.
Our first title is the language learning game described on this page: a game where the practice is the conversation, and a model that understands what you said is the instructor. It is being built by a small team, in close contact with early learners rather than by guessing what they want.
The name is a statement of intent. A waymaker builds the path instead of waiting for one — which is also a fair description of learning to speak a new language.
Closed beta places are limited and go to people who will actually use the product and tell us what breaks. Write to us with the language you are learning and the platform you use, and we will add you to the list.
Partnerships, startup programs and press: the same address. We reply within two business days.