Can You Understand English Recorded on a Potato?
We ran a state-of-the-art speech recognizer over 33,664 real recordings. It fully cleared 60%. The other 40% — the audio you skip on YouTube — is where listening is actually learned.
Method, data, and what it actually takes to reach fluency.
We ran a state-of-the-art speech recognizer over 33,664 real recordings. It fully cleared 60%. The other 40% — the audio you skip on YouTube — is where listening is actually learned.
diff-match-patch was written to reconcile machine text. We point it at human ears — it grades the game, powers the ads, and mails your misses back to you.
We invented an avatar that does not exist yet. If you type for a living — and everyone does now — it might be you.
Yeah. Thought so. Fix it — you won't like it — thank us six months later, if you do the program.
A warning label and a hypothesis: your brain only bothers to understand once the volume of syllables makes not-understanding too expensive.
You know that it works, but you still hate it. We made a few changes so you hate it less: two new exercises, playable today as open previews.
You met one awkward sentence out of thirty-five thousand and you're thinking of leaving. The arithmetic says otherwise: fluency is a volume problem, and volume forgives what polish cannot buy.
Humanity spent twenty years recording itself so machines could learn to listen. The recordings are also the best ear-training material ever assembled for people — the machines never ran out of volunteers; language learners did.
Twelve new languages in one evening — Basque, Galician, Georgian, Esperanto among them — because the volunteers of Common Voice already did the hard part, and letting their work sit unused would be a waste of other people's love.
You know what an LLM is. It got that way by exposure to a large language corpus — not grammar rules, not streak badges. If your knowledge is only going to be passive for now, make it good passive.
Somewhere on your device is a folder of language-learning PDFs you have never opened twice. Gathering material is the learner's favorite way to avoid learning — and the content problem is already solved.
You claim you speak ten languages. A conversation lets you bluff — context, cues, body language. A dictation stream of random native-speed sentences does not. Prove it statistically; I published my own tables first.
You cannot study your way to returning a 200 km/h serve. The return exists only as a reflex built from standing in the box while aces go by. Native speech is the serve.
The Stats tabs read like an enterprise back office — transactions, positions, performance. That is deliberate. Your practice history is a portfolio, and you are its clerk.
Get good at dictation and the AP grammar section becomes almost boring — the patterns are already internalized, the way natives know grammar before it has names.
English is spoken everywhere, better than ever — so why learn anything else? Because English is the reserve language, the way the dollar is the reserve currency. You can transact in it anywhere. You can't belong in it anywhere.
In March 2020 I asked my Uber driver to turn around at the airport door. Then, in traffic, I asked him to turn around again. That 180 across the highway is why this product exists.
I sold my car in 2021 and put the money in index funds. When the office called me back, I spent $800 a month on rideshares — and learned German in the back seat.
Not the two hours a grammar book asks for. Twenty-five minutes — the length of an Uber ride. It's in your pocket now.
Nobody you know would transcribe thousands of machine-learning mp3s to learn a language. That's a fact about your town, not about you. Kevin Kelly's 1,000 true fans, inverted — a letter addressed to the fan.
We built a desktop app because we believed cloud hosting was expensive. Elastic Beanstalk taught us that in 2022. It wasn't a general truth — it was one architecture. Here's what we changed and why the browser is now the primary way to use SiteDictation.
Engineers keep trying to make AI more deterministic. They have the problem backwards. The volume slider in the web app wasn't asked for — it appeared. That's the useful mode.
Better listening comprehension lets you teach your native language in your native language. Here's why dictation practice is the missing piece for language learners — and why it matters for teachers too.
A free French shareware build is live: 1,000 Common Voice sentences, bundled English translations, 100% offline. The full 70,000-sentence corpus is built too.
A reader found that the Translate button silently did nothing in the free Romanian build. Here's what broke, why it shipped that way, and how we verified the fix instead of guessing.
SiteDictation now runs entirely in the browser at sitedictation.com/play — no install, works on phones. Here's why we built it.
The audio is not unclear — the words are unfamiliar. Understanding why that distinction matters, and what the ear is actually being trained to do with imperfect recordings.
Clozemaster builds vocabulary through mass exposure and spaced repetition. It does not build listening. Those are not the same skill, and the difference does not show up until you are standing in front of a native speaker.
The EdTech industry optimizes for how learning feels. This app optimizes for whether it works. Those are not the same thing, and the difference shows up in the data.
The only honest measure of language fluency is your success rate on audio you have never heard before. Here is 14 months of real data, from zero to C1 in Romanian.