We ran three machine ears over 248,222 real recordings in five languages. Sixty-four percent of our French corpus defeats every one of them. Starting today, the game deals accordingly.
Last week we put a state-of-the-art speech recognizer through our English corpus and found that 40% of real recordings — the potato-mic tier you skip on YouTube — defeated it. This week we finished the census for five more languages, and we used three ears instead of one: Whisper's tiny, base, and small models, which form a natural ladder from worst listener to best. Every clip gets a band: the smallest ear that can transcribe it nearly perfectly. Band 1 means even the worst machine ear clears it. Band 4 means none of them do — pure potato.
The census results:
| corpus | clips | band 1 (easiest) | band 4 (potato) |
|---|---|---|---|
| Spanish | 23,763 | 28.9% | 29.6% |
| Portuguese | 99,999 | 21.7% | 37.4% |
| German | 27,146 | 19.1% | 36.6% |
| Italian | 27,062 | 10.5% | 46.7% |
| French | 70,252 | 11.4% | 64.1% |
Read that French row again. Nearly two-thirds of real French recordings — native speakers, ordinary sentences — defeat every machine listener we own. If you have ever felt that French disappears into a blur of liaison and swallowed syllables, you were not imagining it, and it was never just you. The machines agree.
As of today, the dealer in these five languages knows the bands. A new player gets dealt band-1 clips first — audio even the worst machine ear can parse — and the frontier moves up only as your own results prove you ready. You are never told any of this is happening; there is no placement test, no level selector, no congratulations screen when you graduate. The clips just fit. A small share of every session stays wild — drawn from anywhere, potato included — because the whole method runs on meeting audio you cannot handle yet, and we will not build you a comfortable aquarium. Purists can switch the whole system off with one checkbox and take the corpus raw, the way it was for everyone until this week.
We publish how the ranking works because there is nothing to hide: the machine ladder is an instrument, not a secret. What no one can copy is what comes next — calibrating that ladder against the thing the machines don't have: five years of human attempts, miss by miss, in the ledger.