Under the Hood
How does modern language learning actually work, under the hood? This is where we open up Taalhammer: the algorithms that pick your next sentence and schedule your reviews, the speech technology behind every audio card, and the learning science these choices rest on. Each article shows how we tested something and what the numbers said, including the experiments that didn't work.
Looking for practical study advice instead? Start with Language Learning. For the product view of how we use AI, see AI that actually helps you learn a language.
Start here
-
How FSRS Works: The Spaced Repetition Algorithm Explained
FSRS, the spaced repetition algorithm in Anki: its memory model, 21 parameters, how it is trained on review logs, and how it compares with SM-2.
-
How Pronunciation Assessment Works (and How Accurate It Is)
Inside AI pronunciation checkers: speech models, phonetic annotation, GOP scoring, Azure vs Speechace vs ELSA, and how closely machines agree with human raters.
-
Which TTS Voices Get Pronunciation Right? We Measured 12
We measured 12 TTS voices (Azure, Google Chirp 3 HD, ElevenLabs, OpenAI) on minimal pairs. English voices carry 92–97% of contrasts; Norwegian about 80%.
-
Building the first true n+1 algorithm for language learning
Comprehensible input says every sentence should add exactly one new thing. Nobody has built a system that can prove it. This is our attempt, the feature extractor behind it, and the numbers it currently produces.