How fast do native speakers talk? Speech rates in 44 languages
In native videos, Tamil, Japanese, Turkish, Portuguese, Italian and Spanish speakers say about 6 syllables per second, English and German speakers about 4.7. Measured on 16,293 clips in 44 languages, with the speed by CEFR level, by dialect, and against the classic lab study.
Chief Language Geek @ Reelang
In short videos made by native speakers, the fastest languages run at about 6 syllables per second: Tamil (6.20), Japanese (6.19), Turkish (6.14), Portuguese (6.02), Italian (6.00), Greek (5.99), Telugu (5.99), Spanish (5.92) and Russian (5.90) are so close that their order is within the margin of error. The slowest are Vietnamese (4.22), Thai (4.59), English (4.70) and German (4.76), so the median Spanish clip has about a quarter more syllables per second than the median English clip.
I measured this on 16,293 native YouTube clips in 44 languages, sampled from 230,754 clips (9,850 hours, about 41,000 channels) in the Reelang catalog. Most published speech rates come from people reading sentences aloud in a lab. These are vlogs, sketches, recipes, explainers and news clips, the kind of speech a learner actually runs into.
Nine languages share the top
Each bar in the chart is the median clip of a language. The thin line through it shows the middle half of the clips, from the 25th to the 75th percentile. Tap or hover a language to see its sample size.
The top nine sit within 0.3 syllables per second of each other, and their confidence ranges overlap. Tamil's median is somewhere between 6.04 and 6.49 with 95% confidence, Spanish's between 5.82 and 6.02. So read the top of the chart as a fast group, not a podium. The bottom is similar: Norwegian, German, Swedish, Tagalog, Belarusian and English are all between 4.70 and 4.78, with Thai and Vietnamese below them.
The spread inside one language is bigger than the gap between most languages. A clip at the 75th percentile of English (5.16 syllables per second) is as fast as a clip at the 25th percentile of Spanish (5.17). Which speaker you hear matters more than which language they speak.
Japanese needs a note. Japanese speakers and teachers usually count in morae, not syllables. Tōkyō is two syllables (tō-kyō) but four morae (to-o-kyo-o). I counted syllables, as the lab study below did, so that Japanese can be compared with other languages. Counted in morae, the median Japanese clip runs at 7.45 per second, and the middle half of clips at 6.04 to 8.95.
This high school sketch is at the fast end of Japanese: a student gives excuses for being late, in Kansai dialect, and the clip runs at 7.84 syllables per second. The seven seconds below hold six lines.
Words per minute gets it backwards
The most common way to talk about speed is words per minute. Switch the chart above to words per minute and the ranking nearly turns over. Mandarin, Vietnamese and Thai move to the top with 303, 253 and 232 words per minute. Kannada (122), Korean (132), Tamil (137) and Telugu (140) drop to the bottom. Across the 44 languages, the rank correlation between the two measures is -0.25: slightly negative.
The reason is that a word is a different thing in each language. In Mandarin transcripts each character counts as a word, and a character is one syllable. A Korean "word" is everything between two spaces, usually a word plus its particles. A Tamil or Turkish word often carries several endings. So the median English word has 1.31 syllables and the median Tamil word 2.65. English speakers in these clips say 210 words and Tamil speakers 137 words per minute, but Tamil speakers say more syllables. Words per minute is fine for comparing two clips in the same language. Between languages, it mostly measures how long the words are.
Syllables are the closest thing to a common unit, but they do not carry the same amount of meaning everywhere either. In a well-known lab study, Pellegrino, Coupé and Marsico (2011) found that languages which put less information into each syllable are spoken with more syllables per second. Japanese and Spanish were fast and low in information per syllable, Vietnamese and Mandarin slow and high. A later study of 17 languages found that the amount of information per second comes out similar across languages, at around 39 bits per second (Coupé et al., 2019). My ranking fits that pattern: Japanese is at the top, Vietnamese at the bottom and Mandarin in the lower half.
Compared with the lab
Pellegrino, Coupé and Marsico recorded 59 speakers of seven languages, each reading the same 20 short texts translated into their language, and used Vietnamese translations of the texts as a reference. They measured syllables per second after removing every silence longer than 150 ms. All eight languages in their table are in my data:
| Language | Lab, read aloud (2011) | Native videos (2026) | Videos as % of lab |
|---|---|---|---|
| Japanese | 7.84 | 6.19 | 79% |
| Spanish | 7.82 | 5.92 | 76% |
| French | 7.18 | 5.12 | 71% |
| Italian | 6.99 | 6.00 | 86% |
| English | 6.19 | 4.70 | 76% |
| German | 5.97 | 4.76 | 80% |
| Vietnamese | 5.22 | 4.22 | 81% |
| Mandarin | 5.18 | 5.05 | 97% |
The videos run at 71% to 86% of the lab values, with Mandarin as the exception. The order is similar: the rank correlation between the two columns is 0.76 across the 8 languages. Most of the gap probably comes from how the clock runs. The lab removed all pauses over 150 ms. I only skip the gaps between sentences, so the short pauses inside a sentence, the "uh" and the breath before a punchline, count as speaking time. Real speech has a lot of them.
French has the largest gap. The lab counted syllables "as uttered in careful speech", which in French can keep some of the e's that everyday speech drops. I count them as French speakers usually say them: petite is two syllables, not three. Mandarin is the other outlier. Its clips come from a single transcription engine (see the method section), which reads about 3% faster than the others on the same channels. That explains only part of the gap, and I do not have a good explanation for the rest.
Beginner clips are slower, and then speed stops changing
Every clip in the catalog has a CEFR level (A1 to C2). A model assigns it by reading the transcript, sometimes corrected by a human reviewer. The model never sees the timing, so level and speed are two separate measurements.
Clips graded A1 are clearly slower. In the 17 languages with enough A1 clips, the median A1 clip runs at 71% of the speed of B1 clips. Korean A1 is at 57% of B1, French at 65%, English at 74% and Spanish at 90%.
From A2 to C1, the speed barely changes. Across languages, the gap between the fastest and the slowest of the four levels is a median of 11% of the level median. In Spanish it is 1.7%: A2 clips run at 5.87 syllables per second, B1 at 5.88, B2 at 5.97 and C1 at 5.93. In German it is 3%, in English 3.7%. C1 is not reliably faster than B1 either. In Spanish and Dutch it is slightly faster (5.93 against 5.88, 5.27 against 5.15), in English slightly slower (4.60 against 4.75), and in French, Japanese and Russian clearly slower (4.89 against 5.34, 6.16 against 6.85, 4.99 against 6.07). C1 clips include a lot of news, lectures and explainers, which are dense but calm.
So above A1, a higher level means harder words and grammar, not faster speech. The two Spanish clips below show the gap. First, a Colombian clip graded A1: short lines, pauses between them, 4.73 syllables per second in the part below. Then a Rioplatense couple arguing about a dessert, graded A2 for its simple words, at 7.69.
Fast speakers pause less
Speed has two parts: how fast people say the words, and how often they stop. I measured the second part with two numbers from the transcripts' word timings. The mean length of run is the average number of words between two pauses, where a pause is a silence of 250 ms or more (the threshold that worked best in De Jong and Bosker (2013), a study of second language fluency). The phonation ratio is the share of time in which someone is actually speaking.
Within a language, fast clips have longer runs. The rank correlation between a clip's syllables per second and its mean length of run is 0.82 (median over 34 languages), and 0.66 with the phonation ratio. Part of this is built into the measure, because my clock includes short pauses inside a sentence, so read it as a plain description: fast clips pause less.
The pauses are where beginner clips differ most. Spanish A1 clips are only 10% slower than B1 clips in syllables, but their runs are much shorter: 5.6 words between pauses at A1 against 9.6 at B1. In English it is 4.0 against 10.8, in German 3.6 against 8.9. English A1 speakers are speaking 41% of the time, B1 speakers 79%. For a learner, that silence is the time to process what was just said.
Between languages, these numbers are harder to compare, because a word is longer in some languages than in others. Greek (6.6 words per run) and Turkish (6.9) have short runs of long words, Dutch (10.4) and French (9.9) long runs of short words. Multiply by the syllables per word and all four come out at roughly 13 to 16 syllables per run.
Dialects: real gaps, smaller than their reputation
Where a dialect had at least 60 measured clips, I report it separately. To keep the comparison fair, these medians are adjusted to the same mix of levels as the whole language, so a dialect with many A1 clips does not look slow for that reason. Pick any two below.
Spanish
Clips from Spain (6.05 syllables per second), Rioplatense clips from Argentina and Uruguay (6.03) and Colombian clips (6.03) are faster than Mexican clips (5.75) and Andean clips (5.55). Words per minute does not show this at all: Spain, Rioplatense and Mexican clips are all at 198 or 199 words per minute. The confidence ranges of Spain and Mexico just touch. Rioplatense and Mexican do not overlap.
Portuguese
Brazilian Portuguese runs at 6.06 syllables per second, European Portuguese at 5.56. European Portuguese has a reputation for being fast and hard to follow, but in syllables per second it is the slower one here. The reputation more likely comes from the sound: European speakers often reduce or drop unstressed vowels, so words sound compressed. My count works from the written words, so it still counts those vowels. If anything, that makes European Portuguese look faster here than it is. The European sample is also small (80 clips), and its confidence range overlaps the Brazilian one.
The two voiceovers below, one about fitting a phone screen protector and one a recipe, are close in speed by my count: 6.12 syllables per second in the Brazilian part, 6.39 in the European one. Listen for how many vowels you can hear in each.
French
Clips from France run at 5.29, clips from Québec at 4.90, and the confidence ranges do not overlap.
Japanese
Kansai speech has a reputation for being fast. The raw numbers seem to agree (6.28 against 6.16 for standard Japanese), but once the level mix is the same, Kansai comes out at 6.13 and standard Japanese at 6.21. So the raw difference comes from the levels the Kansai clips are graded at, not from the dialect, and both confidence ranges are wide (5.77 to 6.81 and 5.64 to 6.86).
Arabic
Egyptian Arabic is the fastest variety at 6.12, ahead of Moroccan (5.68), Modern Standard Arabic (5.56), Levantine (5.53) and Gulf Arabic (5.37). Tunisian is the slowest at 4.66. The Arabic counts are estimates, because short vowels are not written (see the method section).
German
Standard German runs at 5.00, Swabian at 4.63, Austrian at 4.46, Bavarian at 4.41 and Swiss German at 3.98. The Swiss German transcripts are written in Standard German spelling, so what I count are the syllables of the Standard German words.
English and Mandarin
American, British, Australian and Canadian English all fall between 4.69 and 4.75. Nigerian English is slower at 4.30. Mainland Mandarin runs at 5.27, Taiwanese Mandarin at 4.80, from a smaller sample of 65 Taiwanese clips.
What this means if you are learning a language
The differences between languages are real but modest. A Spanish learner and an English learner face speech that differs by about a quarter in syllables per second, and in both languages the middle half of clips spans more than that. What changes much more is the step from beginner material to everything else. Once you leave A1, clips do not get slower as the level goes down: A2 clips are about as fast as C1 clips. The practical way through is to start with clips that have long pauses and short lines, and then to choose clips by topic and speaker rather than wait for a level where people speak slower.
If you want to try this, the Spanish, Portuguese, French, English and Japanese pages list native clips by level, with the same word-by-word subtitles as the clips on this page. Speed is one half of what makes listening hard. The other half is vocabulary, which I measured in how many words you need to understand native videos. And Christian Lauer's guide for beginners covers the first months, when A1 clips are what you need.
How we measured this
The clips
All clips were in the Reelang catalog on 29 September 2026: publicly shown, with speech, native content and not language lessons, not dubbed, in one language, and not rated unusable by a human reviewer. Almost all are from YouTube. From these 230,754 clips I drew a stratified sample by language, dialect and CEFR level: up to 60 clips per group, chosen by a fixed hash so the sample can be rebuilt. I used the first 30 sentences of each clip. Songs were removed (1,617 sampled clips), and no channel could contribute more than 3 clips to one group. Each clip was then weighted by the size of its group in the catalog, so the numbers describe the catalog, not the sample. All figures are weighted medians, with the 25th and 75th percentiles as the "middle half". Confidence ranges come from 300 bootstrap resamples.
Which languages are ranked
A language needs at least 150 clips and 40 channels in the catalog, and at least 75 clips left after all filters. 44 languages pass. Esperanto passes the numbers but is left out, because it has almost no native speakers. Dialects need at least 100 clips and 20 channels in the catalog and 60 measured clips.
Syllables per second
For each sentence I take the time at which its first word starts and the time at which its last word starts, and divide the syllables of every word except the last by that span. Word start times are the one timing that is accurate in every transcript source; end times are partly estimated by some of the engines. Per clip, the syllables and spans of all sentences are added up. A clip needs at least 5 usable sentences and 10 seconds of speech. The span includes short pauses inside a sentence but not the gaps between sentences. Words per minute uses the same span.
Counting syllables
The syllables come from the text, as each word is normally pronounced:
- For 30 languages, including English, Spanish, French, German, Italian, Portuguese, Russian, Turkish and the Indian languages, each word goes through the pronunciation engine espeak-ng, and every vowel sound or syllabic consonant counts as one syllable.
- Croatian, Serbian, Bosnian, Macedonian, Czech and Slovak are counted with letter rules that include syllabic r and l (Hrvatska has three syllables). Tagalog counts vowel letters.
- Mandarin counts one syllable per Chinese character, Korean one per Hangul block, Vietnamese one per written syllable.
- Japanese is converted to kana with pykakasi and then counted in syllables and in morae.
- Thai uses the dictionary syllable splitter of PyThaiNLP, which slightly undercounts.
- Arabic and Hebrew do not write short vowels, so their syllables are estimated from letters and words with a formula fitted on 25 hand-counted sentences each (about 9% and 12% error per sentence, under 1% overall).
Sentences with digits were dropped, because numbers are read differently in each language. On 102 random sentences in 12 languages, the automatic count matched a hand count exactly in 98, and the totals were equal (1,267 syllables each). The hand counts were made with the automatic count visible, so this is a check, not a blind test. Arabic, Hebrew, Persian and Thai are marked as estimates.
Transcripts
The transcripts are machine-made, by three engines. For clips from the same channel, one engine gives rates about 0.18 syllables per second (3%) higher than the others. For Mandarin, Arabic and Hindi, one engine clearly lost words, so its clips are left out for those languages. Mandarin is measured on 293 clips from a single engine.
Pauses
Mean length of run and phonation ratio need the end time of each word, which two of the engines estimate the same way. I only use clips from those two engines for these measures, and I do not compare them across languages whose words are characters or clusters of characters.
What this sample is not
The catalog is chosen for language learners. It leans toward short clips with clear speech and broad appeal: sketches, vlogs, cooking, explainers. It is not a random sample of how people talk. The unit is the clip, not the speaker, because the transcripts do not separate speakers reliably. Syllables are counted from the text, so when a fast speaker swallows a syllable that the transcript still has, the clip counts it anyway, and fast clips may be slightly overstated. The CEFR level is a model's reading of the transcript.
Use this data
You are welcome to reuse the charts and numbers on this page in articles, teaching materials and research. Please link to this page as the source.
| # | Language | Syllables/s | Middle half of clips | Words/min | Syllables per word | Clips | Hours | Channels | Measured |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Tamil | 6.20 | 5.35 to 7.18 | 137 | 2.65 | 1,619 | 45 | 257 | 219 |
| 2 | Japanese | 6.19 | 5.17 to 7.43 | 224 | 1.71 | 2,837 | 150 | 868 | 441 |
| 3 | Turkish | 6.14 | 5.16 to 7.10 | 161 | 2.29 | 3,849 | 144 | 1,011 | 376 |
| 4 | Portuguese | 6.02 | 5.12 to 6.86 | 202 | 1.81 | 14,642 | 652 | 2,278 | 435 |
| 5 | Italian | 6.00 | 5.39 to 6.59 | 186 | 1.91 | 9,180 | 294 | 1,226 | 631 |
| 6 | Greek | 5.99 | 5.25 to 6.60 | 178 | 2.01 | 3,770 | 124 | 484 | 237 |
| 7 | Telugu | 5.99 | 5.38 to 6.58 | 140 | 2.62 | 2,390 | 85 | 640 | 177 |
| 8 | Spanish | 5.92 | 5.17 to 6.55 | 196 | 1.76 | 30,969 | 1,501 | 3,111 | 1,467 |
| 9 | Russian | 5.90 | 4.93 to 6.77 | 175 | 2.02 | 3,611 | 83 | 984 | 350 |
| 10 | Malay | 5.77 | 5.11 to 6.30 | 152 | 2.10 | 1,570 | 55 | 343 | 211 |
| 11 | Kannada | 5.74 | 5.08 to 6.32 | 122 | 2.74 | 1,668 | 74 | 508 | 216 |
| 12 | Finnish | 5.65 | 4.89 to 6.28 | 154 | 2.16 | 3,550 | 106 | 349 | 247 |
| 13 | Arabic * | 5.63 | 4.75 to 6.52 | 158 | 2.14 | 26,232 | 778 | 3,206 | 1,350 |
| 14 | Indonesian | 5.62 | 4.81 to 6.39 | 153 | 2.17 | 4,249 | 131 | 1,465 | 479 |
| 15 | Macedonian | 5.60 | 4.73 to 6.06 | 162 | 2.03 | 369 | 13 | 75 | 80 |
| 16 | Korean | 5.56 | 4.70 to 6.40 | 132 | 2.51 | 10,246 | 730 | 2,558 | 650 |
| 17 | Romanian | 5.41 | 4.82 to 6.06 | 195 | 1.60 | 3,898 | 176 | 606 | 465 |
| 18 | Bosnian | 5.40 | 4.78 to 5.90 | 159 | 1.98 | 200 | 12 | 63 | 94 |
| 19 | Croatian | 5.40 | 4.91 to 6.07 | 166 | 1.92 | 1,319 | 44 | 326 | 147 |
| 20 | Marathi | 5.36 | 4.79 to 5.92 | 149 | 2.12 | 2,066 | 74 | 446 | 201 |
| 21 | Serbian | 5.36 | 4.79 to 5.84 | 170 | 1.89 | 2,267 | 80 | 256 | 209 |
| 22 | Hungarian | 5.35 | 4.72 to 6.04 | 155 | 2.02 | 6,680 | 435 | 1,369 | 394 |
| 23 | Catalan | 5.28 | 4.52 to 5.78 | 181 | 1.68 | 200 | 9 | 59 | 81 |
| 24 | Czech | 5.28 | 4.62 to 5.93 | 171 | 1.79 | 2,611 | 96 | 337 | 248 |
| 25 | Bulgarian | 5.26 | 4.64 to 5.80 | 155 | 2.02 | 1,564 | 93 | 296 | 254 |
| 26 | Danish | 5.25 | 4.65 to 5.95 | 214 | 1.47 | 1,250 | 78 | 205 | 162 |
| 27 | Bengali | 5.21 | 4.42 to 5.87 | 141 | 2.17 | 2,711 | 236 | 1,276 | 435 |
| 28 | Polish | 5.21 | 4.62 to 5.84 | 181 | 1.75 | 4,129 | 165 | 882 | 288 |
| 29 | Slovak | 5.21 | 4.39 to 5.73 | 173 | 1.76 | 974 | 39 | 135 | 170 |
| 30 | French | 5.12 | 4.59 to 5.82 | 214 | 1.41 | 6,897 | 322 | 1,194 | 621 |
| 31 | Ukrainian | 5.10 | 4.41 to 5.78 | 143 | 2.12 | 1,312 | 67 | 228 | 248 |
| 32 | Hebrew * | 5.07 | 4.33 to 5.86 | 155 | 1.97 | 5,596 | 211 | 732 | 292 |
| 33 | Mandarin Chinese | 5.05 | 4.74 to 5.54 | 303 | 1.00 | 7,146 | 340 | 1,381 | 293 |
| 34 | Dutch | 5.03 | 4.49 to 5.58 | 214 | 1.39 | 7,201 | 281 | 712 | 496 |
| 35 | Persian * | 4.99 | 4.43 to 5.29 | 156 | 1.93 | 748 | 29 | 167 | 105 |
| 36 | Hindi | 4.81 | 4.05 to 5.48 | 175 | 1.61 | 2,690 | 142 | 1,080 | 302 |
| 37 | Norwegian | 4.78 | 4.27 to 5.45 | 199 | 1.46 | 522 | 20 | 120 | 121 |
| 38 | German | 4.76 | 4.13 to 5.44 | 179 | 1.55 | 12,865 | 670 | 2,285 | 875 |
| 39 | Swedish | 4.75 | 4.26 to 5.20 | 196 | 1.45 | 3,061 | 141 | 436 | 358 |
| 40 | Tagalog | 4.75 | 4.29 to 5.23 | 148 | 1.92 | 1,225 | 43 | 471 | 335 |
| 41 | Belarusian | 4.73 | 3.48 to 5.41 | 131 | 2.08 | 157 | 5 | 78 | 93 |
| 42 | English | 4.70 | 4.21 to 5.16 | 210 | 1.31 | 28,732 | 1,002 | 6,201 | 1,060 |
| 43 | Thai * | 4.59 | 4.12 to 5.11 | 232 | 1.19 | 983 | 51 | 285 | 119 |
| 44 | Vietnamese | 4.22 | 3.80 to 4.79 | 253 | 1.00 | 999 | 23 | 332 | 261 |
Syllables/s: the median clip, in syllables per second of speech. Middle half of clips: the 25th to the 75th percentile. Words/min: the median clip, where a "word" is whatever the transcript separates (one character in Mandarin), so this column is not comparable between languages. Clips, hours and channels: the native clips in the catalog. Measured: the clips in the sample the rates are computed from. * Estimated syllable counts (Arabic and Hebrew from a letter formula, Persian and Thai with a known small undercount).
Sources
- Coupé, C., Oh, Y. M., Dediu, D., & Pellegrino, F. (2019). Different languages, similar encoding efficiency: Comparable information rates across the human communicative niche. Science Advances, 5(9), eaaw2594. doi.org/10.1126/sciadv.aaw2594, free full text
- De Jong, N. H., & Bosker, H. R. (2013). Choosing a threshold for silent pauses to measure second language fluency. Proceedings of Disfluency in Spontaneous Speech (DiSS 2013), 17-20. ISCA Archive
- Pellegrino, F., Coupé, C., & Marsico, E. (2011). A cross-language perspective on speech information rate. Language, 87(3), 539-558. doi.org/10.1353/lan.2011.0057, open archive copy
- Tools: espeak-ng 1.52, pykakasi, PyThaiNLP.
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
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.
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.