Mike from Clozemaster asked the right question: what do you use after Duolingo? His answer was Clozemaster. He was right. Now here is the follow-up question nobody is asking: what do you use after Clozemaster? Or rather — what do you use alongside it, for the skill it does not train?
Clozemaster deserves real credit. It took the Tatoeba corpus — hundreds of thousands of real sentences across dozens of languages — and built a spaced repetition system around it at a time when most language apps were still selling gamified vocabulary lists of 500 words. Mass exposure to real sentences, sentence-level practice rather than isolated vocabulary, binary correct/incorrect grading, a corpus large enough that you never run out of material. These are the right instincts, and they influenced this app directly.
The session model also contains a hidden insight. If you set Clozemaster to 50 exercises with 40 reviews and 10 new, you have implicitly created a 20% new-material injection rate — what this app calls the fiat factor. Clozemaster arrived at it through product intuition. Site Dictation makes it explicit, tunable per language and session length, and validated against four years of real transaction data. The student acknowledging the teacher.
Clozemaster presents sentences with gaps. You fill in the missing word. Audio is available — as a supplement. The primary signal is text. You are reading and recalling, with occasional audio to confirm. After thousands of Clozemaster reps you will have broad vocabulary, solid pattern recognition in written form, and a dangerous confidence about spoken conversation.
Dangerous, because it does not survive contact with a native speaker.
"My main problem wasn't simply having to think when I spoke — I was also missing most of what was said. My listening was very bad."
Clozemaster did not cause this. It just did not fix it. You will get to intermediate and find that reading a Romanian newspaper is manageable but following a phone call is not. Native speakers switch to English the moment they notice you are struggling — courteously, but decisively — removing the only real practice opportunity you had. Vocabulary is necessary but not sufficient. The bottleneck is the ear, not the dictionary.
Site Dictation started as a Clozemaster problem. The Tatoeba corpus is real and useful, but the exercises were visual. To train listening you need audio. The first instinct was to synthesize it — feed the sentences to a text-to-speech engine and run dictation against the output.
While searching for a Romanian TTS system, I found something better: Adriana Stan's speech synthesis research at romaniantts.com, openly published, with the underlying recordings exposed. Not the synthesized output — the actual human recordings used to build the system. Real speakers. Real microphones. Real transcripts. The corpus already existed. It just needed to be used differently.
Site Dictation ends after 25 minutes, regardless of how many exercises you got wrong. You are not required to clear the queue. You are not coerced into another hour of Anki-style debt repayment. The only failure is not doing the session. Fixed time, variable output, daily compliance as the only metric that compounds — the Pomodoro principle applied to language acquisition.
The fiat factor mentioned above is tunable: shorter sessions lean on review, longer sessions allow more new material, beginners start at 10% new, advanced learners at 25%. You will fail exercises. That is the point. The failure is information, the repetition handles the rest, and the session ends on time regardless.
The payoff is not a test score. The payoff is that native sources become accessible without preparation.
A Romanian Instagram reel. A conversation overheard on the tram. A podcast in a café. These are not educator-curated inputs adjusted for your level — they are full-speed, ambient, native-register language with background noise, overlapping speech, regional accents, no subtitles, and no patience for learners. This is the actual target environment.
Until listening is trained, these sources are noise. You cannot absorb what you cannot parse. Clozemaster gets you to the edge of this threshold. Dictation on real human audio is what crosses it.
Once you cross it, the input becomes self-reinforcing. Every reel, every overheard sentence, every phone call is a session. You stop depending on apps to provide the practice. The immersion that was previously unavailable — because you lacked the substrate to extract signal from it — becomes permanently available, everywhere, for free.
Clozemaster and Site Dictation are not competing for the same learner at the same stage. Clozemaster builds reading-speed vocabulary. Site Dictation builds listening. A serious learner probably needs both, and the sequence matters: vocabulary first gives the mental hooks, listening work builds the automatic pattern recognition that vocabulary alone never produces.
The question to ask is not which app is better. It is: when I am speaking with a native speaker, what actually fails? If you cannot recall words, Clozemaster has more to offer. If you can recall words but cannot track the conversation at native speed, you have a listening problem. More reading-based practice will not fix it.