How many words do you actually need?

Learners tend to imagine vocabulary as a long uniform slope: 50,000 words in the language, so learning it means climbing 50,000 steps.

It is not a slope. It is a cliff followed by a very long, very gentle plain, and almost all the value is in the cliff.

The distribution

Word frequency follows a steep power-law-ish distribution. In most languages, a small number of words do the overwhelming majority of the work.

Coverage of running text
100% ┤                                    ....────────
     │                        ....────────
 90% ┤              ...───────
     │        ..────
 80% ┤     .──
     │   .─
 70% ┤  ╷
     │  │
 50% ┤ ╷│
     │ ││
     └─┴┴──────────────────────────────────────────────
      100  1k    3k     5k      10k        20k
                  words known

Approximate figures, which vary by language and corpus but hold in shape:

Words known	Coverage of running text
100	~50%
1,000	~75–85%
2,000	~85–90%
3,000	~90–93%
5,000	~95%
8,000–9,000	~98%
20,000	~99%

The first thousand words buy you 80% of the text. The next nineteen thousand buy you the last 19%. Nothing else you can do with your study time comes close to the return on the first thousand.

But coverage is not comprehension

Here is where the numbers mislead if taken alone.

Coverage of 90% sounds excellent. It means one word in ten is unknown — roughly two per sentence. Try reading a page like that: it is exhausting, comprehension is patchy, and you cannot guess from context because the context itself has holes.

The thresholds that matter:

• 95% coverage (~4,000–5,000 families) — reading with a dictionary is possible. One unknown word in twenty. Slow, workable.
• 98% coverage (~8,000–9,000 families) — unassisted reading becomes comfortable. One unknown word in fifty, guessable from context. This is the number that corresponds to what people mean by “I can read the news.”

That gap — from 90% to 98% — is where most of the vocabulary work lives, and it is genuinely long. But it is also the part you do largely through reading rather than through study, because at that level you meet new words in comprehensible contexts.

Speech is much easier than text

Spoken language is far more repetitive than written language. Conversation reuses a small core relentlessly; writing reaches for precision and variety.

Practical consequence: around 1,500–2,000 word families supports real conversation, while the same learner cannot read a novel. This asymmetry is normal and it is why people who “can talk fine” often find a newspaper humiliating.

If your goal is speaking, the target is much closer than the reading figures suggest. Two thousand items, properly active, is a functional speaker.

Word families, not words

Every figure above is in word families — a base word plus its inflections and transparent derivations. Teach, teaches, taught, teaching, teacher is one family.

This matters enormously across languages. A Spanish verb has around fifty inflected forms; a Finnish noun has fifteen cases times two numbers before you start on possessive suffixes. Counting surface forms would make Finnish look ten times harder than English, which is not the experience of learners. Counting families makes the comparison honest.

It also has a study implication: do not make a flashcard for each inflected form. Learn the family and the morphological pattern.

Passive is not active

Every number above describes recognition vocabulary. Your production vocabulary will be considerably smaller — typically half, often less.

That gap is normal, it exists in your first language too, and it is the whole subject of understanding but not speaking. If your goal is speech, the relevant question is not “how many words do I know” but “how many can I retrieve under time pressure”, and the answer is usually humbling.

Producing 1,500 words actively is worth more than recognising 5,000 passively, for almost every practical purpose.

A realistic schedule

At ten new items a day with spaced repetition:

	New items	Time to reach
1,000	10/day	~3.5 months
2,000	10/day	~7 months
3,000	10/day	~10 months
5,000	10/day	~1.5 years
9,000	10/day	~2.5 years

Add review time: the daily load stabilises around 20–40 minutes at ten new cards a day, because mature cards return rarely.

Five a day is the more sustainable number for most working adults, and it still reaches 2,000 within about a year. Twenty a day looks appealing in week one and collapses around month four when the review load peaks. See Anki card design for why the new-cards-per-day setting matters more than any other choice.

Which words

Use a frequency list for the first 1,000–2,000. This is the one place where a pre-made deck is defensible: the top 2,000 words of a language are genuinely universal and there is no advantage to discovering them yourself.

After that, take words from what you read and hear. Frequency lists degrade past a few thousand items, because rank 4,000 in a general corpus may be rank 200 in your actual life. A software engineer, a nurse and a sailor need three different second thousands.

Learn them in sentences. A word learned in isolation is a translation pair; a word learned in a sentence carries its collocations, its grammar and its pronunciation. This is the single largest determinant of whether vocabulary becomes usable, and it is why cards built from your own lessons — sentences you personally reached for and could not produce — outperform any list.

That last workflow is what Teachee automates: the sentence you failed to produce in a lesson arrives already transcribed, translated and phonetically annotated, exports to Anki, and enters the schedule. The frequency list gets you to two thousand. Your own failures get you the rest.