Fluen

Your grammar did not fail. The end of your words did.

Paula La Rosa, Product & Service Designer and Design Engineer

Sector
EdTech · Languages
Role
Product Design · Design Engineering · Research
Year
2026
Stack
React + Vite, TypeScript, Tailwind, GSAP + ScrollTrigger, Three.js, Web Speech API, Web Audio API
Live
fluen-delta.vercel.app/

Context

An English method for Portuguese speakers, built from a real placement assessment instead of a generic program. The thesis is in the name: Fluen, one T short. The dropped final consonant is what breaks plurals and past tense, and it is exactly what the listener never receives.

Problem

People who study English for years still freeze when speaking while reading effortlessly. The usual diagnosis blames grammar, but the assessment showed something else: the structure is there and the end of the word disappears. Without the final consonant, "screens" becomes "screen" and "tested" becomes "test", three pieces of information the listener simply does not get.

Solution

Assessment by skill instead of a single level, which is the average that hides the problem. The target is the Lingua Franca Core: in goes what changes understanding, out goes what only marks an accent. Mistakes come back the same day and only leave the deck after three correct answers in a row in a new context. Writing blocks saving until it meets volume and structure, because writing too little feels just like having studied.

What the research showed

How it was done. A placement test per skill, with reading, writing, speaking and listening measured separately, speech recordings compared word by word and interviews with adults who have studied English for years and freeze in conversation.

Findings

  1. The grammar was there; the end of the word was gone: People who freeze do not get the structure wrong, they drop the final consonant. Plurals and past tense disappear, and the listener loses the information.
  2. A single level hides the gap: The average puts people at a level they do not have in any skill. Reading and listening run ahead of speaking and writing.
  3. Accent is not the problem: People spend time trying to sound American. What breaks the conversation is what changes the meaning, not the accent.

Personas

Fernanda, 38, Project manager at a multinational. Reads reports in English effortlessly, but freezes when she has to speak in meetings with the team abroad. Needs: Speak clearly, even with an accent, and be understood the first time. Gets in the way: Courses that repeat the grammar she already knows.

Lucas, 26, Developer aiming for a job abroad. Studies alone at night on his phone and has no one to practice speaking with. Needs: Hear where he went wrong and practice only what changes understanding. Gets in the way: Apps that score pronunciation without saying what to fix.

Personas composed from the research. Names and ages are illustrative.

What was left out

I discarded the initial plan, focused on aspirated p, t and k and dark L: the placement test showed those were already fine. Item-by-item pronunciation scores were also left out, since speech recognition distorts them.

How to know it worked

  • How many final consonants show up in recorded speech, week over week.
  • How many mistakes leave the deck with three correct answers in a row in a new context.
  • How many days a week have a completed session.

Design decisions

The first hypothesis was wrong, so the method changed

The plan bet on p/t/k aspiration, -ed as a syllable and dark L. The assessment showed those were already fine and that the problem was final consonants, rhythm and /ɜːr/. A method that does not correct itself is not measuring, it is confirming what it already believed.

The goal is not to sound American

Lingua Franca Core (Jenkins, 2000) as the cutoff: in go final consonants, plurals, past tense, rhythm and /ɜːr/. Left out: a perfect th, dark L and schwa. Chasing an accent spends the time that should go to what changes understanding.

Level by skill, not an average

Reading, writing, speaking and listening are assessed separately. A single number hides exactly the gap the method needs to attack, in this case reading and listening ahead of speaking and writing.

Everything on the device, and the exception is stated

No server, no sign-up, no password. The only exception is written on the page itself: in Chrome, speech recognition processes audio on a Google server, and typing avoids it. Hiding that caveat would be easier and less honest.

Results

  • A diagnosis that contradicts the plan: 14 errors grouped into 5 families, the largest being have + participle. The phonetic profile was redone after the assessment.
  • An open engine, no sign-up: The same engine as the study area runs on the landing page: the person says the sentence and compares it word by word before signing in.
  • Evidence treated as evidence: The recognizer corrects what sounds close to a likely word, so the method counts repeated patterns and never an isolated item.