FSRS Spaced Repetition: How to Stop Re-Losing Sentences You Already Unlocked

Say you've worked through the previous six posts. You've used dictation to crack segments you couldn't hear, shadowing to get them into your mouth, and you've checked the stress and intonation. Then you meet the same sentence a week later and it stops you cold, like you'd never heard it.

You didn't practice it wrong. That's just how memory behaves. The variable that matters isn't how hard you worked on it — it's when you met it again.

The interval problem

Come back too early and you spend time on something you already know. Come back too late and you're cracking it from scratch all over again. The efficient moment is right as the memory is about to slip, and that moment is different for every card and every person. Reviewing a sentence you got on the first listen at the same interval as one you missed three times shortchanges both.

What FSRS actually computes

FSRS — Free Spaced Repetition Scheduler — is the algorithm that finds that moment. For every card, it tracks three things:

• Stability — how long this memory holds. It grows each time you recall the card successfully.
• Difficulty — how stubborn this particular card is for you. It climbs when you keep missing it.
• Retrievability — your odds of recalling it at this instant. It decays as time passes.

When you finish a card you answer Again / Hard / Good / Easy, and that answer updates stability and difficulty. The scheduler then computes the date retrievability is projected to fall to 90% and brings the card back that day. Each button shows the interval it would produce before you press it, so you're not grading blind.

Where SM-2 — Anki's default for many years — was closer to "multiply the last interval by a factor," FSRS models the forgetting curve itself and solves backwards for the target probability.

None of that is unique to us; it's how any good spaced-repetition app behaves. The interesting problem is the next one.

Language isn't text

English word on the front, definition on the back. Review that a hundred times and you will absolutely know what the word means. Then you hear it go by in an actual conversation and you still miss it.

Of course you do. What you reviewed was letters. Rehearsing "read it, recall it" makes you better at reading it and recalling it, and your listening was never in the loop. Everything dictation made you notice the hard way — how two words collapse into each other in this exact spot, how far this one reduces once it loses its stress — is not on that card.

Spaced repetition isn't the wrong idea. The thing being repeated is the wrong object.

So we put the sound on the card

In Emergence's desktop app, when you turn a subtitle line into a card, the audio of that line as it was actually spoken gets cut out of the video and attached to the card. Not a synthetic voice reading it back — the original speaker. If the line needs surrounding context to make sense, a context clip of up to 30 seconds is stored alongside it.

That changes what review is. The sentence text starts out blurred while the audio plays. You listen, replay it if you need to, and only then reveal the text to check yourself. Instead of reading and recalling, the review itself is a listening rep.

Word cards work a little differently. Save an unknown word from a subtitle and the whole sentence it came from stays on the card, with the word marked where it appeared. On top of that, an AI-written example sentence is added along with audio of that sentence. That audio is synthesized — it isn't the original speaker's pronunciation — but it gets you hearing the same word in a second context.

You have to say it to move on

For sentence cards, review doesn't end at listening. Say the line back and your pronunciation is assessed — accuracy, fluency, completeness, and prosody each get a score — and the overall score proposes your FSRS grade for you:

• 90 and above → Easy
• 75 and above → Good
• 60 and above → Hard
• below 60 → Again

This matters because in ordinary spaced repetition, "how well did I know that" is pure self-report. Press Good on something you only half-remembered and the card comes back too late, and you'll miss it again. Using the pronunciation score as the grade means the number feeding the scheduler is measured rather than felt. It's still only a suggestion — override it, or skip assessment and grade by hand, whenever you want.

What the screen is for

"Visual learning" usually conjures picture cards. That isn't what the display does here. It's there to show you the evidence for what your ears just did.

• A word isn't stranded next to a dictionary gloss — it's shown in its original subtitle sentence with its position marked.
• The sentence text is hidden by default, which forces you to process the sound before you can see the answer.
• After assessment, each word is color-coded so you can see exactly where the pronunciation went wrong.
• You can overlay the waveform of the original against yours, along with the intonation contour.
• The stats screen shows the distribution of stability, difficulty, and review intervals, plus a 12-week practice heatmap.

The ordering is deliberate: judge by ear first, then look at the evidence. Flip it — show the text first — and the listening rep evaporates into a reading rep.

How to start

1. Turn a segment you unlocked into a card. Use a line you already processed today with dictation and shadowing, not something new you haven't worked through.
2. Listen before you look. Play it a few times blurred, then reveal.
3. Say it out loud. Reproduce what you heard before you check the text.
4. Take the grade the score suggests. Override it only when it clearly disagrees with how it felt.
5. Clear today's queue and go back to input. Review doesn't replace input; it's what makes input stay.

How many cards should you make

Make every card you could make and you'll be buried under a backlog within a week and quit. One or two sentences a day is plenty. It's the same principle as in the massive input post: trying to capture everything you watch burns out a method that only works if it's sustainable. Watch broadly, pull out the one line that trips you, and keep only that line as a card.

That's the end of this course. Watch a lot, hear it precisely, say it back, check the pronunciation and the melody, and then make it stick. Spaced repetition earns its place at the end of that sequence — start stacking cards without the five steps before it, and you'll just be diligently reviewing sentences you still can't hear.