How many words do you need to understand native videos?
With the 5,000 most common English words, the typical native clip is 95.6% words you know. Spanish needs about 6,800 words to reach 95%, and in 13 of 23 languages even 10,000 are not enough. Data from 35,615 native clips, per language and per level.
Chief Language Geek @ Reelang
If you know the 5,000 most common English words, the typical native English video clip is 95.6% words you know, and 58.4% of clips reach 95%, the level at which listening studies find learners understand well. Spanish needs about 6,800 words for the typical clip to reach 95%, and in 13 of the 23 languages I measured, 10,000 words are not enough.
These numbers come from 35,615 native clips in 23 languages from the Reelang catalog, counted word by word against a frequency list for each language. Pick a language and a vocabulary size below to see what it gets you.
The first 1,000 words do most of the work
Word frequency is lopsided. A few hundred words (the, and, is, you, que, de) make up a large share of everything anyone says, and each word further down the list shows up less often. This pattern is known as Zipf's law, and it holds in every language here.
Across the 18 languages I can rank without a caveat, the typical clip is 80.7% covered by the top 1,000 words, 86.7% by the top 2,000 and 92.4% by the top 5,000. The first thousand words get you to four known words in five. The next four thousand add about twelve points.
The thresholds that matter for listening sit inside those last points. Two of them come up again and again in the research on how many words you need:
- 95%, one unknown word in 20. Batia Laufer proposed it for reading in 1989. For listening, van Zeeland and Schmitt (2013) found that most listeners already coped at 90%, and that at 95% second-language listeners understood well with much less variation between them.
- 98%, one unknown word in 50. Hu and Nation (2000) found that learners reading fiction on their own needed about this much for adequate comprehension.
At 150 words a minute, a normal speaking pace, 95% means about seven unknown words every minute (Nation, 2006). 98% means three. You can feel that difference in the demo below. It uses real lines from clips in the study and blanks out every word that is not in the top words you pick.
With 2,000 words, the meatball line is 58.3% known: albóndigas, parmesano, rallado and orégano are not in the list at all, and perejil (parsley) sits at rank 13,520. The line about the cat is 96.2% known at the same 2,000 words. Both clips are Spanish and both are labelled A2.
English and Spanish, in numbers
English. The typical clip reaches 90% at 2,021 words, 95% at 4,462 words and 98% at 9,403 words. With 3,000 words, 24.9% of English clips are at 95% or more. With 5,000 words, 58.4% are, and 15.2% are at 98% or more.
Spanish. The typical clip reaches 90% at 2,754 words and 95% at 6,781 words. It does not reach 98% within 10,000 words. With 3,000 words, 10.8% of Spanish clips are at 95% or more. With 5,000 words, 32.0% are.
So for Spanish, a learner with 3,000 words can find about one clip in ten where 95% of the words are familiar. That is a real pool to learn from, because the words you do not know yet show up surrounded by words you do. This Colombian clip is one of them: 88.4% of its words are in the top 1,000 and 96.3% in the top 2,000.
Knowing the words is not the same as catching them. The speakers in that clip talk at a median of 243 words per minute, much faster than a textbook recording. Vocabulary coverage tells you whether you could understand the transcript. It does not tell you whether you can keep up with the speed.
Flat languages and steep languages
The curves do not all have the same shape. In Danish, Swedish and Dutch, the top 1,000 words already cover 84.4% to 86.5% of a typical clip, and going from 1,000 to 5,000 words adds only 8.3 to 9.3 points. English, French, Spanish, Portuguese and Italian add 11.5 to 12.6 points over the same range. The steepest curves are Indonesian (+20.8), Ukrainian (+18.3), Arabic (+18.1), Russian (+16.5), Finnish (+15.3), Turkish (+15.2) and Polish (+14.3).
A steep curve means the first thousand words leave more of the clip unknown, and every further thousand is worth more. Part of this is the language. Part of it is how words are counted, and that needs explaining before you compare two languages.
The lists here are lemmatized: dije, dijo and decir all count as one word, decir. That is fair to Spanish, which has a lot of verb forms. But a lemmatizer only works when it recognizes the form. In Czech, Finnish, Hungarian, Russian and Ukrainian, 9.0% to 10.8% of the running words in a typical clip are forms the list could not match to any lemma. Many of those are ordinary words with a case ending the tool did not strip. They count as unknown here, so for those five languages the numbers are a floor: real coverage is higher. They are marked with an asterisk in the table below.
Compounds work against German, Dutch and the Nordic languages in a similar way. A German compound like Küchenschrank (kitchen cupboard) counts as unknown unless the whole compound is in the list, even when you know both parts.
The level label barely changes the vocabulary
Every clip in the catalog has a CEFR level, set by a model that reads the transcript. Here they are grouped into five bands, from A1 (A0 counts as A1) to C1 (C2 counts as C1). If the label tracked vocabulary, higher levels would need more words. For the first 2,000 words, they hardly do. Taking the median across languages (A1 has enough clips in 11 of them, C1 in 20), the top 2,000 words cover 84.0% of an A1 clip, 84.0% at A2, 85.2% at B1, 86.1% at B2 and 84.1% at C1.
The level does show at the top of the range. With 5,000 words, you would know 95% of the words in 30.2% of A1 clips, 20.6% of A2 clips, 22.6% of B1 clips, 15.5% of B2 clips and 4.5% of C1 clips. Length is part of the reason: the median C1 clip in the sample has 338 running words, the median A1 clip 81. A short clip has fewer chances to use a rare word.
The clips with the lowest coverage are mostly about food. Of the 25 English clips in the sample with the lowest coverage at 2,000 words, about two thirds are recipes, food reviews or snack tastings, and 21 are labelled A1 or A2. In Spanish, 11 of the lowest 25 are about food, and three are children's songs. The grammar in a recipe is simple ("add the garlic, stir, done"), which is why it gets a low label. The nouns are not. In the English list, garlic is word 6,395 and tablespoon 11,551. In Spanish, ajo is 6,605 and sartén (frying pan) 11,125.
Here are two English clips. The first is a film trailer labelled B2: 96.9% of its words are in the top 2,000, and it passes 95% at 1,300 words. The second is a slow cooker recipe labelled A2: 72.3% of its words are in the top 2,000, and it only passes 95% at 11,551 words, the rank of tablespoon.
The trailer is not easy either, for a different reason: its speakers talk at a median of 215 words per minute. A level label mixes speed, grammar and vocabulary into one letter, and the vocabulary part is what gets lost. If you know between 2,000 and 5,000 words, look at the topic of a clip as well as its level: food, cooking and product talk use many words from far down the list.
Varieties differ too
Arabic is measured with a separate frequency list for each dialect. The other languages use one list for all their varieties. With 2,000 words, the typical Egyptian Arabic clip is 86.1% covered, Iraqi 79.6% and Moroccan 72.5%. For Moroccan, 9.5% of the words in a typical clip are not in its list at all, so part of that gap is the list. In French, Québec clips come out at 90.6% against 87.6% for clips from France. Brazilian Portuguese is at 87.4%, European Portuguese at 80.3%, but only 61 European Portuguese clips passed the filters and they are scored against a different list, so treat that one with care.
How this compares with Nation's estimate
The best-known figure for spoken English comes from Paul Nation's 2006 paper "How large a vocabulary is needed for reading and listening?". Using a corpus of spoken New Zealand English, he found that 3,000 word families plus proper nouns give over 95% coverage, and that 6,000 to 7,000 word families are needed for 98%. Webb and Rodgers (2009) got similar numbers for film scripts: 3,000 families plus proper nouns and marginal words for 95.76%, and 6,000 families plus the same extras for 98.15%.
My numbers for English are higher: 4,462 words for the typical clip to reach 95%, 9,403 for 98%. Taking all English clips together as one long text, 5,000 words cover 95.1%. There are three reasons the numbers differ, and none of them means one study is wrong:
- Families are bigger than lemmas. A word family groups happy, unhappy, happiness and happily as one unit. A lemma list counts them as four. So 3,000 families correspond to a larger number of lemmas.
- The list is built mostly from written text. Everyday spoken words sit lower in it than you would expect: alright is word 3,102, kinda 2,333 and bye 3,273.
- Short videos stick to one topic. A two-minute recipe spends most of its words on one topic's vocabulary, where a long conversation moves around and reuses more general words.
What this means if you are learning a language
The first 2,000 or 3,000 words are worth learning on purpose, with a frequency-ordered deck or course. By that point, a real share of native clips is within reach at 95%, and from there you can pick up many new words from context instead of from a list. Christian Lauer's guide for beginners covers how to get to that point quickly.
After that, choose clips by topic, not only by level. A learner with 3,000 Spanish words can follow the cat sketch above and still get lost in the meatball recipe, although both are labelled A2. And for Czech, Finnish, Hungarian, Russian and Ukrainian, expect the numbers here to feel harder than they are: much of what counts as unknown is a word you know with an ending you have not met yet.
If you want to practise with clips like the ones above, the Spanish, English and German pages list native clips by level, with the same word-by-word subtitles as the clips on this page. More on why understanding comes before speaking: how to stop translating in your head.
How I measured this
The clips. All clips were in the Reelang catalog on 29 September 2026 and in rotation: publicly shown, with speech, native content (not language lessons), not dubbed and in one language. For each language I sampled up to 500 clips per CEFR level and weighted each level by its share of the catalog, so the numbers describe the catalog as a whole. Clips with fewer than 50 words, clips where one caption line repeats (more than 30% of lines), clips faster than 330 words per minute of duration, and clips where more than 35% of the words are not in the language's list (usually a transcript in the wrong language) were removed. After that, 35,615 clips in the 23 ranked languages remained, with 13.8 million running words. The median clip is 81 seconds long. 35,600 of them are from YouTube, 8 from TikTok and 7 from Instagram.
The words. Each transcript is split into words on whitespace and punctuation, lowercased, with digits removed. Each word is looked up in a frequency list for its language, built from wordfreq 3.1.1 (Speer, 2022): the 25,000 most frequent word forms, grouped by lemma with simplemma (other tools for Arabic, Hebrew and the Asian languages) and ranked by the total frequency of each lemma. Forms not in the list are lemmatized with simplemma as a fallback. Proper nouns (capitalized, not at the start of a sentence, not in the list) and marginal words like uh, mhm and haha count as known, as Nation (2006) and Webb and Rodgers (2009) do. German capitalizes every noun, and Hindi and Arabic have no capital letters, so those languages get no proper-noun credit. In English, contraction halves count as the full word (didn't counts as did and not).
The numbers. Coverage is the share of running words in a clip whose lemma is among the top N. The percentages above are medians across clips. For the "clips at 95%" figures, each clip's exact threshold is computed from its own word ranks, so there is no interpolation. The "typical clip reaches 95% at" figure is the smallest vocabulary size at which the median clip passes 95%.
Not ranked. Hebrew prefixes (ve-, ha-, be-, le-) are not split off, so known words count as unknown. Tamil transcripts are spoken Tamil and the list is written Tamil. Japanese, Chinese and Korean are counted in content words only, and Vietnamese in syllables, so their numbers cannot be compared with the rest. A language needs at least 300 clips to be ranked. Norwegian has 449; every other ranked language has more than 1,000. The Turkish list has only 9,088 lemmas, so for Turkish the charts stop at 8,000 words, and "10,000+" means the typical clip does not reach 95% with the whole list.
Known biases. The numbers describe transcripts, not audio, so speech-recognition errors count as words. Knowing the top N words perfectly is an ideal case, and idioms are not counted as units. The CEFR level is a model's judgement of the transcript. And the catalog is chosen for learners: it leans toward clips with clear speech and broad appeal, so it is not a random sample of everything on YouTube or TikTok.
Use this data
You are welcome to reuse the charts and numbers on this page in your own writing, teaching or research. Please link to this page as the source.
| Language | Clips | 1,000 | 2,000 | 5,000 | 95%+ | 98%+ | 95% at |
|---|---|---|---|---|---|---|---|
| Danish | 1,122 | 86.5% | 91.1% | 94.8% | 47.0% | 8.4% | 5,292 |
| English | 2,002 | 83.7% | 89.9% | 95.6% | 58.4% | 15.2% | 4,462 |
| Swedish | 1,325 | 85.6% | 89.8% | 94.0% | 37.6% | 5.6% | 6,768 |
| Dutch | 1,841 | 84.4% | 89.0% | 93.7% | 32.2% | 3.5% | 7,109 |
| French | 1,961 | 82.7% | 88.4% | 94.2% | 40.1% | 5.8% | 5,804 |
| Norwegian | 449 | 83.5% | 88.1% | 93.0% | 25.0% | 3.8% | 9,478 |
| Spanish | 2,195 | 81.6% | 87.3% | 93.5% | 32.0% | 5.4% | 6,781 |
| Portuguese | 1,865 | 81.4% | 87.2% | 94.0% | 38.4% | 7.3% | 6,121 |
| Romanian | 1,602 | 81.4% | 87.0% | 92.5% | 25.3% | 5.8% | 9,586 |
| Esperanto | 1,007 | 80.0% | 86.3% | 92.3% | 28.6% | 4.5% | 10,000+ |
| Italian | 1,880 | 79.7% | 85.6% | 91.9% | 21.2% | 2.7% | 9,269 |
| Turkish | 1,322 | 77.1% | 85.6% | 92.3% | 25.4% | 5.5% | 10,000+ † |
| German | 1,933 | 80.0% | 84.8% | 90.3% | 11.3% | 1.1% | 10,000+ |
| Greek | 1,796 | 78.0% | 83.5% | 89.0% | 6.7% | 1.0% | 10,000+ |
| Hindi | 1,417 | 77.2% | 82.4% | 88.3% | 6.6% | 1.2% | 10,000+ |
| Arabic | 1,916 | 73.1% | 82.2% | 91.2% | 17.8% | 2.7% | 10,000+ |
| Polish | 1,503 | 73.9% | 81.0% | 88.2% | 8.2% | 1.0% | 10,000+ |
| Indonesian | 1,582 | 68.6% | 80.6% | 89.4% | 10.4% | 2.3% | 10,000+ |
| Czech * | 1,239 | 73.4% | 80.0% | 86.8% | 3.9% | 0.5% | 10,000+ |
| Finnish * | 1,565 | 71.7% | 79.3% | 87.0% | 3.6% | 0.8% | 10,000+ |
| Hungarian * | 1,732 | 71.7% | 78.1% | 84.5% | 0.4% | 0.0% | 10,000+ |
| Russian * | 1,305 | 70.0% | 77.8% | 86.5% | 2.9% | 0.1% | 10,000+ |
| Ukrainian * | 1,056 | 65.4% | 73.4% | 83.7% | 0.8% | 0.0% | 10,000+ |
Columns 1,000, 2,000 and 5,000: the median share of running words in a clip covered by the top 1,000, 2,000 or 5,000 lemmas. 95%+ and 98%+: the share of clips in which a learner who knows the top 5,000 lemmas knows at least 95% or 98% of the words. 95% at: the vocabulary size at which the median clip passes 95% (10,000+ means not within 10,000). * Lower bound: 9% or more of the words in a typical clip could not be matched to the list, so real coverage is higher. † The Turkish list has only 9,088 lemmas: the typical Turkish clip does not reach 95% with the whole list.
Sources
- Hu, M. H.-c., & Nation, I. S. P. (2000). Unknown vocabulary density and reading comprehension. Reading in a Foreign Language, 13(1), 403-430. doi.org/10.64152/10125/66973 (free full text)
- Laufer, B. (1989). What percentage of text-lexis is essential for comprehension? In C. Lauren & M. Nordman (Eds.), Special language: From humans thinking to thinking machines (pp. 316-323). Multilingual Matters. Record
- Nation, I. S. P. (2006). How large a vocabulary is needed for reading and listening? The Canadian Modern Language Review, 63(1), 59-82. doi.org/10.3138/cmlr.63.1.59, free PDF
- Speer, R. (2022). wordfreq v3.0 [data and software]. Zenodo. doi.org/10.5281/zenodo.7199437. This study used release 3.1.1 from github.com/rspeer/wordfreq.
- van Zeeland, H., & Schmitt, N. (2013). Lexical coverage in L1 and L2 listening comprehension: The same or different from reading comprehension? Applied Linguistics, 34(4), 457-479. doi.org/10.1093/applin/ams074
- Webb, S., & Rodgers, M. P. H. (2009). The lexical coverage of movies. Applied Linguistics, 30(3), 407-427. doi.org/10.1093/applin/amp010
Start watching
Pick a language and start the feed. Real native speech, graded by how hard it is to follow. No quiz, no signup.
Pick a languageRead next
Comprehensible input for beginners
Comprehensible input only works once you can follow what is being said. Christian Lauer on how to get to that point fast, and how to tell when you are there.
How to stop translating in your head: why comprehension comes before speaking
You stop translating in your head by giving your brain so much comprehensible speech that it stops needing to translate at all. Translation is a workaround your brain builds when it doesn't have enough direct experience with a language. The fix isn't a trick or a mental discipline, it's volume: hours of listening to real speech you mostly understand, until your brain builds a direct path from sound to meaning instead of routing through your first language.
Why comprehensible input works for Mandarin learners
Comprehensible input works for Mandarin because the language's biggest obstacles, tones, character-based literacy, and a wide gap between spoken and written registers, are all things you learn by ear before you learn by rule. Grammar drills and flashcards teach you to recognize Mandarin in isolation. Comprehensible input teaches you to recognize it the way native speakers actually produce it, at natural speed, in context, with the tonal and rhythmic patterns intact. That distinction is the whole article.