You begin a language color blind. Your judgment learns of colors you are being punished for not seeing — and then, even as an adult, you start to see them. Here are the three stages, and the five years of data I tested them against.
At the beginning you try to hear and simply fail — not from inattention, but because you lack the lexical and grammatical structure to receive the speech at all. The sentence arrives as weather. My first year of Romanian dictation looks exactly like this in the ledger: success on never-heard sentences grinding from 19% up to 45% over ten slow months. No spurt. No miracle. Weather, transcribed badly.
Then something that feels like cheating: you begin to pass exercises you did not hear. Grammar and memory answer instead. In Romanian, firme and firmei — "firms" and "of the firm" — differ by a vowel that fast speech nearly swallows; for years I wrote the genitive i not because my ear reported it but because my judgment demanded it. You cannot trust your perception yet; you suspect every clip of being a trap. So reason leads. This is the strange, backwards phase where a second language argues that judgment is superior to intuition — you know things about the sound that you cannot yet hear.
And then, under enough pressure, the synthesis. My hypothesis for why: hearing is cheaper than remembering. By this point your memory is maintaining thousands of sentences under spaced-repetition pressure, and the brain wants to unload. The audio cues it has been punished over are sitting right there — cheaper to extract at listening time than to store and retrieve. So it learns to extract them. The overloaded memory is not the cost of the method; it is the forcing function. You start to hear the richness of native audio — the very endings you used to write on faith — and the grammar that was only remembered becomes something you perceive. The colors arrive. Even in an adult brain.
This would be just a pretty theory, so I replayed all 126,097 of my Romanian attempts through the scheduler and watched the review debt and the cold-start success rate move month by month. The prediction: perception should improve fastest exactly when memory pressure is highest. What the ledger shows: the steepest climb in five years of data — cold-start success from 41.5% to 77.2% in five months — happened precisely as my overdue-review queue exploded from ~500 sentences to over 5,000, and my volume went vertical to two hours a day. And one more thing, which surprised me: at every debt level from five hundred to fifteen thousand overdue sentences, my success on reviews stayed flat between 75% and 85%. The queue never rotted. The debt that classical spaced-repetition doctrine treats as an emergency turned out to be harmless — possibly even the engine.
Honesty about the limits: this is one learner's ledger, and the pressure, the schedule change, and the volume explosion all arrived within eight days of each other, so no single factor can be isolated. But the direction is exactly what the compression story predicts, and the flat review curve quietly humbles every algorithm debate I have ever had with myself: neither strict scheduling nor generous novelty was the secret. Both are shadows of the only variable that visibly mattered — volume. The one true catastrophe in my data wasn't algorithmic at all: in 2023 my progress stopped because I ran out of sentences.
Which leaves the hardest truth for last. In the gym, beginners get their gains first and meet the long feedback-less grind later. Language runs the opposite curve: the grind comes first — stage one and stage two are the grind — and the beginner gains arrive at the end, dressed as a miracle. B2 is a beginner. C1 is intermediate. Ask any barbell. If you quit in the gray months because nothing seemed to be happening, you quit while the colors were being mixed.
Ten months of nothing, then thirty-six points in five. The spurt is backloaded. Survive to it.