That is my count. Every failed Romanian transcription since March 2021, logged. It is not the number of times the method failed me. It is the method.
My app records every attempt: the sentence I heard, what I typed, whether it matched. Since March 11, 2021, I have attempted 126,097 Romanian transcriptions and failed 28,883 of them. Nearly one attempt in four, for five years, wrong. I went from zero Romanian to C1 — living in Boston, not Bucharest — and the ledger says the failures were not the tax on the journey. They were the vehicle.
The lifetime total flatters me; most of it is maintenance. The number that matters is smaller and harder: 41,553 exercises in the first fourteen months — March 2021 through April 2022 — with 11,018 failures inside it. That is what the road from zero to functional fluency cost, at a 26.5% failure rate the entire way. I remember 41,553 the way schoolboys remember 1453. A city fell.
The start was rockier than the average: in my first ninety days I failed a third of everything I touched. It never got comfortable — through my most intense months the daily success rate swung between 47% and 85%. There was no phase where the failing stopped. There was only a phase where I stopped minding.
Here is what a failed transcription actually is: a sentence your ear could not resolve today, scheduled to return. You sleep. Overnight, memory consolidation does work no amount of harder studying can do — and tomorrow the same sentence is slightly less hard. Not solved. Slightly less hard. You have adapted, the way skin calluses: not by avoiding the friction but by meeting it on schedule.
The first failed exercise is a pointer to forty thousand more such failures — on a hard but relatively sure road to fluency.
The algorithm's job is not to protect you from hard sentences. It is to throw enough punches that you become a good boxer, and to time the punches so your nervous system rebuilds between rounds. The consistency requirement is real but humble: show up, fail your quota, sleep. The brain is very resistant to adapting to a language — especially when you live where it is not spoken — and the only argument it accepts is volume, delivered daily, with the misses recycled.
I did not do this in ten minutes a day. In January and February 2022 I practiced on 53 of 57 days, a median of two hours a day, peaking at four and a half. It was a professional hobby — closer to a training block than a habit. What drove it wasn't discipline: on January 3, 2022 I added a timed-sessions counter to my own app, having just read a book about habits, and the counter did the rest. There was no rule. You send the message just by measuring it.
I trusted the method through the failure noise for a simple reason: I had seen spaced repetition work on my Latin, and here I was unleashing it on a living language with troves of native audio. The whole thing ran on localhost with no CSS. The company came later. The 28,883 came first.
My ledger also holds 79,000 German attempts at a 33% failure rate — and my German is nowhere near C1. The difference: alongside Romanian dictation I paid for daily conversation lessons; for German I never did. Dictation built me a broad passive base in German that lessons would convert quickly — but by itself it did not close the deal. This is no mono-therapy. The app is the gym work. You still have to get in the ring.
It ends somewhere strange, and I will be honest about that too. These days, landing at Otopeni, reading Bine ai venit and Ieșire on the signs, I sometimes laugh to myself — the country feels like a level in a game I trained for, and for a moment I am not sure whether I am arriving or just starting another session. That is what tens of thousands of scheduled failures do: the language stops being a subject and becomes terrain. Are you willing to risk that for the sake of a language? Then the first failed exercise is waiting, and it is not an omen. It is the odometer starting to turn.